#include "tu_cs.h"
-#define OVERFLOW_FLAG_REG REG_A6XX_CP_SCRATCH_REG(0)
-
void
tu_bo_list_init(struct tu_bo_list *list)
{
static void
tu_tiling_config_update_tile_layout(struct tu_tiling_config *tiling,
const struct tu_device *dev,
- uint32_t pixels)
+ const struct tu_render_pass *pass)
{
- const uint32_t tile_align_w = 64; /* note: 32 when no input attachments */
- const uint32_t tile_align_h = 16;
+ const uint32_t tile_align_w = pass->tile_align_w;
const uint32_t max_tile_width = 1024;
/* note: don't offset the tiling config by render_area.offset,
.height = 1,
};
tiling->tile0.extent = (VkExtent2D) {
- .width = align(ra_width, tile_align_w),
- .height = align(ra_height, tile_align_h),
+ .width = util_align_npot(ra_width, tile_align_w),
+ .height = align(ra_height, TILE_ALIGN_H),
};
if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN)) {
/* start with 2x2 tiles */
tiling->tile_count.width = 2;
tiling->tile_count.height = 2;
- tiling->tile0.extent.width = align(DIV_ROUND_UP(ra_width, 2), tile_align_w);
- tiling->tile0.extent.height = align(DIV_ROUND_UP(ra_height, 2), tile_align_h);
+ tiling->tile0.extent.width = util_align_npot(DIV_ROUND_UP(ra_width, 2), tile_align_w);
+ tiling->tile0.extent.height = align(DIV_ROUND_UP(ra_height, 2), TILE_ALIGN_H);
}
/* do not exceed max tile width */
while (tiling->tile0.extent.width > max_tile_width) {
tiling->tile_count.width++;
tiling->tile0.extent.width =
- align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
+ util_align_npot(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
}
/* will force to sysmem, don't bother trying to have a valid tile config
* TODO: just skip all GMEM stuff when sysmem is forced?
*/
- if (!pixels)
+ if (!pass->gmem_pixels)
return;
/* do not exceed gmem size */
- while (tiling->tile0.extent.width * tiling->tile0.extent.height > pixels) {
+ while (tiling->tile0.extent.width * tiling->tile0.extent.height > pass->gmem_pixels) {
if (tiling->tile0.extent.width > MAX2(tile_align_w, tiling->tile0.extent.height)) {
tiling->tile_count.width++;
tiling->tile0.extent.width =
- align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
+ util_align_npot(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
} else {
/* if this assert fails then layout is impossible.. */
- assert(tiling->tile0.extent.height > tile_align_h);
+ assert(tiling->tile0.extent.height > TILE_ALIGN_H);
tiling->tile_count.height++;
tiling->tile0.extent.height =
- align(DIV_ROUND_UP(ra_height, tiling->tile_count.height), tile_align_h);
+ align(DIV_ROUND_UP(ra_height, tiling->tile_count.height), TILE_ALIGN_H);
}
}
}
: tile->begin.y + tiling->tile0.extent.height;
}
-enum a3xx_msaa_samples
-tu_msaa_samples(uint32_t samples)
-{
- switch (samples) {
- case 1:
- return MSAA_ONE;
- case 2:
- return MSAA_TWO;
- case 4:
- return MSAA_FOUR;
- case 8:
- return MSAA_EIGHT;
- default:
- assert(!"invalid sample count");
- return MSAA_ONE;
- }
-}
-
-static enum a4xx_index_size
-tu6_index_size(VkIndexType type)
-{
- switch (type) {
- case VK_INDEX_TYPE_UINT16:
- return INDEX4_SIZE_16_BIT;
- case VK_INDEX_TYPE_UINT32:
- return INDEX4_SIZE_32_BIT;
- default:
- unreachable("invalid VkIndexType");
- return INDEX4_SIZE_8_BIT;
- }
-}
-
-unsigned
+void
tu6_emit_event_write(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
- enum vgt_event_type event,
- bool need_seqno)
-{
- unsigned seqno = 0;
+ enum vgt_event_type event)
+{
+ bool need_seqno = false;
+ switch (event) {
+ case CACHE_FLUSH_TS:
+ case WT_DONE_TS:
+ case RB_DONE_TS:
+ case PC_CCU_FLUSH_DEPTH_TS:
+ case PC_CCU_FLUSH_COLOR_TS:
+ case PC_CCU_RESOLVE_TS:
+ need_seqno = true;
+ break;
+ default:
+ break;
+ }
tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
if (need_seqno) {
tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- seqno = ++cmd->scratch_seqno;
- tu_cs_emit(cs, seqno);
+ tu_cs_emit(cs, 0);
}
+}
- return seqno;
+static void
+tu6_emit_flushes(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs,
+ enum tu_cmd_flush_bits flushes)
+{
+ /* Experiments show that invalidating CCU while it still has data in it
+ * doesn't work, so make sure to always flush before invalidating in case
+ * any data remains that hasn't yet been made available through a barrier.
+ * However it does seem to work for UCHE.
+ */
+ if (flushes & (TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR))
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_COLOR_TS);
+ if (flushes & (TU_CMD_FLAG_CCU_FLUSH_DEPTH |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH))
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_DEPTH_TS);
+ if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_COLOR)
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_COLOR);
+ if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_DEPTH)
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_DEPTH);
+ if (flushes & TU_CMD_FLAG_CACHE_FLUSH)
+ tu6_emit_event_write(cmd_buffer, cs, CACHE_FLUSH_TS);
+ if (flushes & TU_CMD_FLAG_CACHE_INVALIDATE)
+ tu6_emit_event_write(cmd_buffer, cs, CACHE_INVALIDATE);
+ if (flushes & TU_CMD_FLAG_WFI)
+ tu_cs_emit_wfi(cs);
}
+/* "Normal" cache flushes, that don't require any special handling */
+
static void
-tu6_emit_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu_emit_cache_flush(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs)
{
- tu6_emit_event_write(cmd, cs, 0x31, false);
+ tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.cache.flush_bits);
+ cmd_buffer->state.cache.flush_bits = 0;
}
-static void
-tu6_emit_lrz_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+/* Renderpass cache flushes */
+
+void
+tu_emit_cache_flush_renderpass(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs)
{
- tu6_emit_event_write(cmd, cs, LRZ_FLUSH, false);
+ tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.renderpass_cache.flush_bits);
+ cmd_buffer->state.renderpass_cache.flush_bits = 0;
}
-static void
-tu6_emit_wfi(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+/* Cache flushes for things that use the color/depth read/write path (i.e.
+ * blits and draws). This deals with changing CCU state as well as the usual
+ * cache flushing.
+ */
+
+void
+tu_emit_cache_flush_ccu(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs,
+ enum tu_cmd_ccu_state ccu_state)
{
- if (cmd->wait_for_idle) {
- tu_cs_emit_wfi(cs);
- cmd->wait_for_idle = false;
+ enum tu_cmd_flush_bits flushes = cmd_buffer->state.cache.flush_bits;
+
+ assert(ccu_state != TU_CMD_CCU_UNKNOWN);
+
+ /* Changing CCU state must involve invalidating the CCU. In sysmem mode,
+ * the CCU may also contain data that we haven't flushed out yet, so we
+ * also need to flush. Also, in order to program RB_CCU_CNTL, we need to
+ * emit a WFI as it isn't pipelined.
+ */
+ if (ccu_state != cmd_buffer->state.ccu_state) {
+ if (cmd_buffer->state.ccu_state != TU_CMD_CCU_GMEM) {
+ flushes |=
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH;
+ cmd_buffer->state.cache.pending_flush_bits &= ~(
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH);
+ }
+ flushes |=
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH |
+ TU_CMD_FLAG_WFI;
+ cmd_buffer->state.cache.pending_flush_bits &= ~(
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH);
+ }
+
+ tu6_emit_flushes(cmd_buffer, cs, flushes);
+ cmd_buffer->state.cache.flush_bits = 0;
+
+ if (ccu_state != cmd_buffer->state.ccu_state) {
+ struct tu_physical_device *phys_dev = cmd_buffer->device->physical_device;
+ tu_cs_emit_regs(cs,
+ A6XX_RB_CCU_CNTL(.offset =
+ ccu_state == TU_CMD_CCU_GMEM ?
+ phys_dev->ccu_offset_gmem :
+ phys_dev->ccu_offset_bypass,
+ .gmem = ccu_state == TU_CMD_CCU_GMEM));
+ cmd_buffer->state.ccu_state = ccu_state;
}
}
}
const struct tu_image_view *iview = fb->attachments[a].attachment;
- enum a6xx_depth_format fmt = tu6_pipe2depth(cmd->state.pass->attachments[a].format);
+ const struct tu_render_pass_attachment *attachment =
+ &cmd->state.pass->attachments[a];
+ enum a6xx_depth_format fmt = tu6_pipe2depth(attachment->format);
tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_BUFFER_INFO, 6);
tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = fmt).value);
tu_cs_image_ref(cs, iview, 0);
- tu_cs_emit(cs, cmd->state.pass->attachments[a].gmem_offset);
+ tu_cs_emit(cs, attachment->gmem_offset);
tu_cs_emit_regs(cs,
A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = fmt));
A6XX_GRAS_LRZ_BUFFER_PITCH(0),
A6XX_GRAS_LRZ_FAST_CLEAR_BUFFER_BASE(0));
- tu_cs_emit_regs(cs,
- A6XX_RB_STENCIL_INFO(0));
-
- /* enable zs? */
+ if (attachment->format == VK_FORMAT_S8_UINT) {
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_INFO, 6);
+ tu_cs_emit(cs, A6XX_RB_STENCIL_INFO(.separate_stencil = true).value);
+ tu_cs_image_ref(cs, iview, 0);
+ tu_cs_emit(cs, attachment->gmem_offset);
+ } else {
+ tu_cs_emit_regs(cs,
+ A6XX_RB_STENCIL_INFO(0));
+ }
}
static void
tu_cs_emit_regs(cs,
A6XX_SP_SRGB_CNTL(.dword = subpass->srgb_cntl));
- tu_cs_emit_regs(cs,
- A6XX_RB_RENDER_COMPONENTS(.dword = subpass->render_components));
- tu_cs_emit_regs(cs,
- A6XX_SP_FS_RENDER_COMPONENTS(.dword = subpass->render_components));
-
tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(fb->layers - 1));
}
A6XX_SP_TP_WINDOW_OFFSET(.x = x1, .y = y1));
}
+static void
+tu_cs_emit_draw_state(struct tu_cs *cs, uint32_t id, struct tu_draw_state state)
+{
+ uint32_t enable_mask;
+ switch (id) {
+ case TU_DRAW_STATE_PROGRAM:
+ case TU_DRAW_STATE_VI:
+ case TU_DRAW_STATE_FS_CONST:
+ /* The blob seems to not enable this (DESC_SETS_LOAD) for binning, even
+ * when resources would actually be used in the binning shader.
+ * Presumably the overhead of prefetching the resources isn't
+ * worth it.
+ */
+ case TU_DRAW_STATE_DESC_SETS_LOAD:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM |
+ CP_SET_DRAW_STATE__0_SYSMEM;
+ break;
+ case TU_DRAW_STATE_PROGRAM_BINNING:
+ case TU_DRAW_STATE_VI_BINNING:
+ enable_mask = CP_SET_DRAW_STATE__0_BINNING;
+ break;
+ case TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM;
+ break;
+ case TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM:
+ enable_mask = CP_SET_DRAW_STATE__0_SYSMEM;
+ break;
+ default:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM |
+ CP_SET_DRAW_STATE__0_SYSMEM |
+ CP_SET_DRAW_STATE__0_BINNING;
+ break;
+ }
+
+ tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(state.size) |
+ enable_mask |
+ CP_SET_DRAW_STATE__0_GROUP_ID(id) |
+ COND(!state.size, CP_SET_DRAW_STATE__0_DISABLE));
+ tu_cs_emit_qw(cs, state.iova);
+}
+
+/* note: get rid of this eventually */
+static void
+tu_cs_emit_sds_ib(struct tu_cs *cs, uint32_t id, struct tu_cs_entry entry)
+{
+ tu_cs_emit_draw_state(cs, id, (struct tu_draw_state) {
+ .iova = entry.size ? entry.bo->iova + entry.offset : 0,
+ .size = entry.size / 4,
+ });
+}
+
static bool
use_hw_binning(struct tu_cmd_buffer *cmd)
{
const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ /* XFB commands are emitted for BINNING || SYSMEM, which makes it incompatible
+ * with non-hw binning GMEM rendering. this is required because some of the
+ * XFB commands need to only be executed once
+ */
+ if (cmd->state.xfb_used)
+ return true;
+
if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
return false;
if (cmd->state.framebuffer->layers > 1)
return true;
+ if (cmd->has_tess)
+ return true;
+
return cmd->state.tiling_config.force_sysmem;
}
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
+ tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5, 7);
+ tu_cs_emit(cs, cmd->state.tiling_config.pipe_sizes[tile->pipe] |
+ CP_SET_BIN_DATA5_0_VSC_N(tile->slot));
+ tu_cs_emit_qw(cs, cmd->vsc_draw_strm.iova + tile->pipe * cmd->vsc_draw_strm_pitch);
+ tu_cs_emit_qw(cs, cmd->vsc_draw_strm.iova + (tile->pipe * 4) + (32 * cmd->vsc_draw_strm_pitch));
+ tu_cs_emit_qw(cs, cmd->vsc_prim_strm.iova + (tile->pipe * cmd->vsc_prim_strm_pitch));
- tu_cs_reserve(cs, 3 + 11);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(11));
-
- /* if (no overflow) */ {
- tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5, 7);
- tu_cs_emit(cs, cmd->state.tiling_config.pipe_sizes[tile->pipe] |
- CP_SET_BIN_DATA5_0_VSC_N(tile->slot));
- tu_cs_emit_qw(cs, cmd->vsc_data.iova + tile->pipe * cmd->vsc_data_pitch);
- tu_cs_emit_qw(cs, cmd->vsc_data.iova + (tile->pipe * 4) + (32 * cmd->vsc_data_pitch));
- tu_cs_emit_qw(cs, cmd->vsc_data2.iova + (tile->pipe * cmd->vsc_data2_pitch));
-
- tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
- tu_cs_emit(cs, 0x0);
-
- /* use a NOP packet to skip over the 'else' side: */
- tu_cs_emit_pkt7(cs, CP_NOP, 2);
- } /* else */ {
- tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
- tu_cs_emit(cs, 0x1);
- }
+ tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
+ tu_cs_emit(cs, 0x0);
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
-
- tu_cs_emit_regs(cs,
- A6XX_RB_UNKNOWN_8804(0));
-
- tu_cs_emit_regs(cs,
- A6XX_SP_TP_UNKNOWN_B304(0));
-
- tu_cs_emit_regs(cs,
- A6XX_GRAS_UNKNOWN_80A4(0));
} else {
tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
tu_cs_emit(cs, 0x1);
tu_resolve_sysmem(cmd, cs, src, dst, fb->layers, &cmd->state.tiling_config.render_area);
}
+static void
+tu6_emit_sysmem_resolves(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ const struct tu_subpass *subpass)
+{
+ if (subpass->resolve_attachments) {
+ /* From the documentation for vkCmdNextSubpass, section 7.4 "Render Pass
+ * Commands":
+ *
+ * End-of-subpass multisample resolves are treated as color
+ * attachment writes for the purposes of synchronization. That is,
+ * they are considered to execute in the
+ * VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT pipeline stage and
+ * their writes are synchronized with
+ * VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT. Synchronization between
+ * rendering within a subpass and any resolve operations at the end
+ * of the subpass occurs automatically, without need for explicit
+ * dependencies or pipeline barriers. However, if the resolve
+ * attachment is also used in a different subpass, an explicit
+ * dependency is needed.
+ *
+ * We use the CP_BLIT path for sysmem resolves, which is really a
+ * transfer command, so we have to manually flush similar to the gmem
+ * resolve case. However, a flush afterwards isn't needed because of the
+ * last sentence and the fact that we're in sysmem mode.
+ */
+ tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS);
+ tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
+
+ /* Wait for the flushes to land before using the 2D engine */
+ tu_cs_emit_wfi(cs);
+
+ for (unsigned i = 0; i < subpass->color_count; i++) {
+ uint32_t a = subpass->resolve_attachments[i].attachment;
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ tu6_emit_sysmem_resolve(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
+ }
+ }
+}
+
static void
tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
}
}
-static void
-tu6_emit_restart_index(struct tu_cs *cs, uint32_t restart_index)
-{
- tu_cs_emit_regs(cs,
- A6XX_PC_RESTART_INDEX(restart_index));
-}
-
static void
tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
const struct tu_physical_device *phys_dev = cmd->device->physical_device;
- tu6_emit_cache_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 0xfffff);
tu_cs_emit_regs(cs,
A6XX_RB_CCU_CNTL(.offset = phys_dev->ccu_offset_bypass));
+ cmd->state.ccu_state = TU_CMD_CCU_SYSMEM;
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E04, 0x00100000);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE04, 0x8);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE00, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9101, 0xffff00);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9107, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9236, 1);
+ tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9236,
+ A6XX_VPC_UNKNOWN_9236_POINT_COORD_INVERT(0));
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9300, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_SO_OVERRIDE,
A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9801, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9806, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9980, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9990, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9981, 0x3);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9E72, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9108, 0x3);
- tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B304, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B309, 0x000000a2);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8804, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A4, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A5, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A6, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8805, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8806, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8878, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8879, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
tu_cs_emit_write_reg(cs, REG_A6XX_VFD_MODE_CNTL, 0x00000000);
tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
- /* Set not to use streamout by default, */
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, 0);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, 0);
-
tu_cs_emit_regs(cs,
A6XX_SP_HS_CTRL_REG0(0));
tu_cs_sanity_check(cs);
}
-static void
-tu6_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- unsigned seqno;
-
- seqno = tu6_emit_event_write(cmd, cs, RB_DONE_TS, true);
-
- tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
- CP_WAIT_REG_MEM_0_POLL_MEMORY);
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(seqno));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
-
- seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
-
- tu_cs_emit_pkt7(cs, CP_WAIT_MEM_GTE, 4);
- tu_cs_emit(cs, CP_WAIT_MEM_GTE_0_RESERVED(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- tu_cs_emit(cs, CP_WAIT_MEM_GTE_3_REF(seqno));
-}
-
static void
update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
tu_cs_emit_regs(cs,
A6XX_VSC_BIN_SIZE(.width = tiling->tile0.extent.width,
.height = tiling->tile0.extent.height),
- A6XX_VSC_SIZE_ADDRESS(.bo = &cmd->vsc_data,
- .bo_offset = 32 * cmd->vsc_data_pitch));
+ A6XX_VSC_DRAW_STRM_SIZE_ADDRESS(.bo = &cmd->vsc_draw_strm,
+ .bo_offset = 32 * cmd->vsc_draw_strm_pitch));
tu_cs_emit_regs(cs,
A6XX_VSC_BIN_COUNT(.nx = tiling->tile_count.width,
tu_cs_emit(cs, tiling->pipe_config[i]);
tu_cs_emit_regs(cs,
- A6XX_VSC_PIPE_DATA2_ADDRESS(.bo = &cmd->vsc_data2),
- A6XX_VSC_PIPE_DATA2_PITCH(cmd->vsc_data2_pitch),
- A6XX_VSC_PIPE_DATA2_ARRAY_PITCH(cmd->vsc_data2.size));
+ A6XX_VSC_PRIM_STRM_ADDRESS(.bo = &cmd->vsc_prim_strm),
+ A6XX_VSC_PRIM_STRM_PITCH(cmd->vsc_prim_strm_pitch),
+ A6XX_VSC_PRIM_STRM_LIMIT(cmd->vsc_prim_strm_pitch - 64));
tu_cs_emit_regs(cs,
- A6XX_VSC_PIPE_DATA_ADDRESS(.bo = &cmd->vsc_data),
- A6XX_VSC_PIPE_DATA_PITCH(cmd->vsc_data_pitch),
- A6XX_VSC_PIPE_DATA_ARRAY_PITCH(cmd->vsc_data.size));
+ A6XX_VSC_DRAW_STRM_ADDRESS(.bo = &cmd->vsc_draw_strm),
+ A6XX_VSC_DRAW_STRM_PITCH(cmd->vsc_draw_strm_pitch),
+ A6XX_VSC_DRAW_STRM_LIMIT(cmd->vsc_draw_strm_pitch - 64));
}
static void
tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
CP_COND_WRITE5_0_WRITE_MEMORY);
- tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE_REG(i)));
+ tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_DRAW_STRM_SIZE_REG(i)));
tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
- tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_draw_strm_pitch - 64));
tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
- tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(1 + cmd->vsc_data_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(1 + cmd->vsc_draw_strm_pitch));
tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
CP_COND_WRITE5_0_WRITE_MEMORY);
- tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE2_REG(i)));
+ tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_PRIM_STRM_SIZE_REG(i)));
tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
- tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data2_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_prim_strm_pitch - 64));
tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
- tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(3 + cmd->vsc_data2_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(3 + cmd->vsc_prim_strm_pitch));
}
tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
- tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
-
- tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
- tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(OVERFLOW_FLAG_REG) |
- CP_MEM_TO_REG_0_CNT(1 - 1));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
-
- /*
- * This is a bit awkward, we really want a way to invert the
- * CP_REG_TEST/CP_COND_REG_EXEC logic, so that we can conditionally
- * execute cmds to use hwbinning when a bit is *not* set. This
- * dance is to invert OVERFLOW_FLAG_REG
- *
- * A CP_NOP packet is used to skip executing the 'else' clause
- * if (b0 set)..
- */
-
- /* b0 will be set if VSC_DATA or VSC_DATA2 overflow: */
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
-
- tu_cs_reserve(cs, 3 + 7);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(7));
-
- /* if (b0 set) */ {
- /*
- * On overflow, mirror the value to control->vsc_overflow
- * which CPU is checking to detect overflow (see
- * check_vsc_overflow())
- */
- tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
- tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(OVERFLOW_FLAG_REG) |
- CP_REG_TO_MEM_0_CNT(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_overflow));
-
- tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
- tu_cs_emit(cs, 0x0);
-
- tu_cs_emit_pkt7(cs, CP_NOP, 2); /* skip 'else' when 'if' is taken */
- } /* else */ {
- tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
- tu_cs_emit(cs, 0x1);
- }
}
static void
tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
tu_cs_emit(cs, UNK_2D);
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
- tu6_cache_flush(cmd, cs);
+ /* This flush is probably required because the VSC, which produces the
+ * visibility stream, is a client of UCHE, whereas the CP needs to read the
+ * visibility stream (without caching) to do draw skipping. The
+ * WFI+WAIT_FOR_ME combination guarantees that the binning commands
+ * submitted are finished before reading the VSC regs (in
+ * emit_vsc_overflow_test) or the VSC_DATA buffer directly (implicitly as
+ * part of draws).
+ */
+ tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS);
tu_cs_emit_wfi(cs);
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
+}
+
+static void
+tu_emit_input_attachments(struct tu_cmd_buffer *cmd,
+ const struct tu_subpass *subpass,
+ struct tu_cs_entry *ib,
+ bool gmem)
+{
+ /* note: we can probably emit input attachments just once for the whole
+ * renderpass, this would avoid emitting both sysmem/gmem versions
+ *
+ * emit two texture descriptors for each input, as a workaround for
+ * d24s8, which can be sampled as both float (depth) and integer (stencil)
+ * tu_shader lowers uint input attachment loads to use the 2nd descriptor
+ * in the pair
+ * TODO: a smarter workaround
+ */
+
+ if (!subpass->input_count)
+ return;
+
+ struct tu_cs_memory texture;
+ VkResult result = tu_cs_alloc(&cmd->sub_cs, subpass->input_count * 2,
+ A6XX_TEX_CONST_DWORDS, &texture);
+ assert(result == VK_SUCCESS);
+
+ for (unsigned i = 0; i < subpass->input_count * 2; i++) {
+ uint32_t a = subpass->input_attachments[i / 2].attachment;
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ struct tu_image_view *iview =
+ cmd->state.framebuffer->attachments[a].attachment;
+ const struct tu_render_pass_attachment *att =
+ &cmd->state.pass->attachments[a];
+ uint32_t *dst = &texture.map[A6XX_TEX_CONST_DWORDS * i];
+
+ memcpy(dst, iview->descriptor, A6XX_TEX_CONST_DWORDS * 4);
+
+ if (i % 2 == 1 && att->format == VK_FORMAT_D24_UNORM_S8_UINT) {
+ /* note this works because spec says fb and input attachments
+ * must use identity swizzle
+ */
+ dst[0] &= ~(A6XX_TEX_CONST_0_FMT__MASK |
+ A6XX_TEX_CONST_0_SWIZ_X__MASK | A6XX_TEX_CONST_0_SWIZ_Y__MASK |
+ A6XX_TEX_CONST_0_SWIZ_Z__MASK | A6XX_TEX_CONST_0_SWIZ_W__MASK);
+ dst[0] |= A6XX_TEX_CONST_0_FMT(FMT6_S8Z24_UINT) |
+ A6XX_TEX_CONST_0_SWIZ_X(A6XX_TEX_Y) |
+ A6XX_TEX_CONST_0_SWIZ_Y(A6XX_TEX_ZERO) |
+ A6XX_TEX_CONST_0_SWIZ_Z(A6XX_TEX_ZERO) |
+ A6XX_TEX_CONST_0_SWIZ_W(A6XX_TEX_ONE);
+ }
+
+ if (!gmem)
+ continue;
+
+ /* patched for gmem */
+ dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
+ dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
+ dst[2] =
+ A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
+ A6XX_TEX_CONST_2_PITCH(cmd->state.tiling_config.tile0.extent.width * att->cpp);
+ dst[3] = 0;
+ dst[4] = cmd->device->physical_device->gmem_base + att->gmem_offset;
+ dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
+ for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
+ dst[i] = 0;
+ }
+
+ struct tu_cs cs;
+ tu_cs_begin_sub_stream(&cmd->sub_cs, 9, &cs);
+
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_FRAG, 3);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_FS_TEX) |
+ CP_LOAD_STATE6_0_NUM_UNIT(subpass->input_count * 2));
+ tu_cs_emit_qw(&cs, texture.iova);
+
+ tu_cs_emit_pkt4(&cs, REG_A6XX_SP_FS_TEX_CONST_LO, 2);
+ tu_cs_emit_qw(&cs, texture.iova);
- cmd->wait_for_idle = false;
+ tu_cs_emit_regs(&cs, A6XX_SP_FS_TEX_COUNT(subpass->input_count * 2));
+
+ *ib = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
+}
+
+static void
+tu_set_input_attachments(struct tu_cmd_buffer *cmd, const struct tu_subpass *subpass)
+{
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_emit_input_attachments(cmd, subpass, &cmd->state.ia_gmem_ib, true);
+ tu_emit_input_attachments(cmd, subpass, &cmd->state.ia_sysmem_ib, false);
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 6);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM, cmd->state.ia_gmem_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM, cmd->state.ia_sysmem_ib);
}
static void
-tu_emit_load_clear(struct tu_cmd_buffer *cmd,
- const VkRenderPassBeginInfo *info)
+tu_emit_renderpass_begin(struct tu_cmd_buffer *cmd,
+ const VkRenderPassBeginInfo *info)
{
struct tu_cs *cs = &cmd->draw_cs;
tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
const struct VkRect2D *renderArea)
{
- const struct tu_physical_device *phys_dev = cmd->device->physical_device;
const struct tu_framebuffer *fb = cmd->state.framebuffer;
assert(fb->width > 0 && fb->height > 0);
tu6_emit_bin_size(cs, 0, 0, 0xc00000); /* 0xc00000 = BYPASS? */
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BYPASS));
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_COLOR, false);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_DEPTH, false);
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
-
- tu6_emit_wfi(cmd, cs);
- tu_cs_emit_regs(cs,
- A6XX_RB_CCU_CNTL(.offset = phys_dev->ccu_offset_bypass));
+ tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_SYSMEM);
/* enable stream-out, with sysmem there is only one pass: */
tu_cs_emit_regs(cs,
/* Do any resolves of the last subpass. These are handled in the
* tile_store_ib in the gmem path.
*/
- const struct tu_subpass *subpass = cmd->state.subpass;
- if (subpass->resolve_attachments) {
- for (unsigned i = 0; i < subpass->color_count; i++) {
- uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED)
- tu6_emit_sysmem_resolve(cmd, cs, a,
- subpass->color_attachments[i].attachment);
- }
- }
+ tu6_emit_sysmem_resolves(cmd, cs, cmd->state.subpass);
tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- tu6_emit_lrz_flush(cmd, cs);
-
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
tu_cs_sanity_check(cs);
}
-
static void
tu6_tile_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
struct tu_physical_device *phys_dev = cmd->device->physical_device;
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
/* lrz clear? */
- tu6_emit_cache_flush(cmd, cs);
-
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- /* TODO: flushing with barriers instead of blindly always flushing */
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_COLOR, false);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_DEPTH, false);
-
- tu_cs_emit_wfi(cs);
- tu_cs_emit_regs(cs,
- A6XX_RB_CCU_CNTL(.offset = phys_dev->ccu_offset_gmem, .gmem = 1));
+ tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_GMEM);
const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
if (use_hw_binning(cmd)) {
tu6_emit_tile_select(cmd, cs, tile);
tu_cs_emit_call(cs, &cmd->draw_cs);
- cmd->wait_for_idle = true;
if (use_hw_binning(cmd)) {
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
-
- tu_cs_reserve(cs, 3 + 2);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(2));
-
- /* if (no overflow) */ {
- tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
- tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
- }
+ tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
}
tu_cs_emit_ib(cs, &cmd->state.tile_store_ib);
tu_cs_emit_regs(cs,
A6XX_GRAS_LRZ_CNTL(0));
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
- tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
+ tu6_emit_event_write(cmd, cs, PC_CCU_RESOLVE_TS);
tu_cs_sanity_check(cs);
}
{
const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ if (use_hw_binning(cmd))
+ cmd->use_vsc_data = true;
+
tu6_tile_render_begin(cmd, &cmd->cs);
for (uint32_t y = 0; y < tiling->tile_count.height; y++) {
tu6_sysmem_render_begin(cmd, &cmd->cs, &tiling->render_area);
tu_cs_emit_call(&cmd->cs, &cmd->draw_cs);
- cmd->wait_for_idle = true;
tu6_sysmem_render_end(cmd, &cmd->cs);
}
tiling->render_area = *render_area;
tiling->force_sysmem = false;
- tu_tiling_config_update_tile_layout(tiling, dev, cmd->state.pass->gmem_pixels);
+ tu_tiling_config_update_tile_layout(tiling, dev, cmd->state.pass);
tu_tiling_config_update_pipe_layout(tiling, dev);
tu_tiling_config_update_pipes(tiling, dev);
}
-const struct tu_dynamic_state default_dynamic_state = {
- .viewport =
- {
- .count = 0,
- },
- .scissor =
- {
- .count = 0,
- },
- .line_width = 1.0f,
- .depth_bias =
- {
- .bias = 0.0f,
- .clamp = 0.0f,
- .slope = 0.0f,
- },
- .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
- .depth_bounds =
- {
- .min = 0.0f,
- .max = 1.0f,
- },
- .stencil_compare_mask =
- {
- .front = ~0u,
- .back = ~0u,
- },
- .stencil_write_mask =
- {
- .front = ~0u,
- .back = ~0u,
- },
- .stencil_reference =
- {
- .front = 0u,
- .back = 0u,
- },
-};
-
-static void UNUSED /* FINISHME */
-tu_bind_dynamic_state(struct tu_cmd_buffer *cmd_buffer,
- const struct tu_dynamic_state *src)
-{
- struct tu_dynamic_state *dest = &cmd_buffer->state.dynamic;
- uint32_t copy_mask = src->mask;
- uint32_t dest_mask = 0;
-
- tu_use_args(cmd_buffer); /* FINISHME */
-
- /* Make sure to copy the number of viewports/scissors because they can
- * only be specified at pipeline creation time.
- */
- dest->viewport.count = src->viewport.count;
- dest->scissor.count = src->scissor.count;
- dest->discard_rectangle.count = src->discard_rectangle.count;
-
- if (copy_mask & TU_DYNAMIC_VIEWPORT) {
- if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
- src->viewport.count * sizeof(VkViewport))) {
- typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
- src->viewport.count);
- dest_mask |= TU_DYNAMIC_VIEWPORT;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_SCISSOR) {
- if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
- src->scissor.count * sizeof(VkRect2D))) {
- typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
- src->scissor.count);
- dest_mask |= TU_DYNAMIC_SCISSOR;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_LINE_WIDTH) {
- if (dest->line_width != src->line_width) {
- dest->line_width = src->line_width;
- dest_mask |= TU_DYNAMIC_LINE_WIDTH;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DEPTH_BIAS) {
- if (memcmp(&dest->depth_bias, &src->depth_bias,
- sizeof(src->depth_bias))) {
- dest->depth_bias = src->depth_bias;
- dest_mask |= TU_DYNAMIC_DEPTH_BIAS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_BLEND_CONSTANTS) {
- if (memcmp(&dest->blend_constants, &src->blend_constants,
- sizeof(src->blend_constants))) {
- typed_memcpy(dest->blend_constants, src->blend_constants, 4);
- dest_mask |= TU_DYNAMIC_BLEND_CONSTANTS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DEPTH_BOUNDS) {
- if (memcmp(&dest->depth_bounds, &src->depth_bounds,
- sizeof(src->depth_bounds))) {
- dest->depth_bounds = src->depth_bounds;
- dest_mask |= TU_DYNAMIC_DEPTH_BOUNDS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_COMPARE_MASK) {
- if (memcmp(&dest->stencil_compare_mask, &src->stencil_compare_mask,
- sizeof(src->stencil_compare_mask))) {
- dest->stencil_compare_mask = src->stencil_compare_mask;
- dest_mask |= TU_DYNAMIC_STENCIL_COMPARE_MASK;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_WRITE_MASK) {
- if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
- sizeof(src->stencil_write_mask))) {
- dest->stencil_write_mask = src->stencil_write_mask;
- dest_mask |= TU_DYNAMIC_STENCIL_WRITE_MASK;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_REFERENCE) {
- if (memcmp(&dest->stencil_reference, &src->stencil_reference,
- sizeof(src->stencil_reference))) {
- dest->stencil_reference = src->stencil_reference;
- dest_mask |= TU_DYNAMIC_STENCIL_REFERENCE;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DISCARD_RECTANGLE) {
- if (memcmp(&dest->discard_rectangle.rectangles,
- &src->discard_rectangle.rectangles,
- src->discard_rectangle.count * sizeof(VkRect2D))) {
- typed_memcpy(dest->discard_rectangle.rectangles,
- src->discard_rectangle.rectangles,
- src->discard_rectangle.count);
- dest_mask |= TU_DYNAMIC_DISCARD_RECTANGLE;
- }
- }
-}
-
static VkResult
tu_create_cmd_buffer(struct tu_device *device,
struct tu_cmd_pool *pool,
goto fail_scratch_bo;
/* TODO: resize on overflow */
- cmd_buffer->vsc_data_pitch = device->vsc_data_pitch;
- cmd_buffer->vsc_data2_pitch = device->vsc_data2_pitch;
- cmd_buffer->vsc_data = device->vsc_data;
- cmd_buffer->vsc_data2 = device->vsc_data2;
+ cmd_buffer->vsc_draw_strm_pitch = device->vsc_draw_strm_pitch;
+ cmd_buffer->vsc_prim_strm_pitch = device->vsc_prim_strm_pitch;
+ cmd_buffer->vsc_draw_strm = device->vsc_draw_strm;
+ cmd_buffer->vsc_prim_strm = device->vsc_prim_strm;
return VK_SUCCESS;
list_del(&cmd_buffer->pool_link);
- for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++)
- free(cmd_buffer->descriptors[i].push_set.set.mapped_ptr);
-
tu_cs_finish(&cmd_buffer->cs);
tu_cs_finish(&cmd_buffer->draw_cs);
tu_cs_finish(&cmd_buffer->draw_epilogue_cs);
static VkResult
tu_reset_cmd_buffer(struct tu_cmd_buffer *cmd_buffer)
{
- cmd_buffer->wait_for_idle = true;
-
cmd_buffer->record_result = VK_SUCCESS;
tu_bo_list_reset(&cmd_buffer->bo_list);
tu_cs_reset(&cmd_buffer->draw_epilogue_cs);
tu_cs_reset(&cmd_buffer->sub_cs);
- for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
- cmd_buffer->descriptors[i].valid = 0;
- cmd_buffer->descriptors[i].push_dirty = false;
- }
+ for (unsigned i = 0; i < MAX_BIND_POINTS; i++)
+ memset(&cmd_buffer->descriptors[i].sets, 0, sizeof(cmd_buffer->descriptors[i].sets));
cmd_buffer->status = TU_CMD_BUFFER_STATUS_INITIAL;
return tu_reset_cmd_buffer(cmd_buffer);
}
+/* Initialize the cache, assuming all necessary flushes have happened but *not*
+ * invalidations.
+ */
+static void
+tu_cache_init(struct tu_cache_state *cache)
+{
+ cache->flush_bits = 0;
+ cache->pending_flush_bits = TU_CMD_FLAG_ALL_INVALIDATE;
+}
+
VkResult
tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
const VkCommandBufferBeginInfo *pBeginInfo)
}
memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
+ cmd_buffer->state.index_size = 0xff; /* dirty restart index */
+
+ tu_cache_init(&cmd_buffer->state.cache);
+ tu_cache_init(&cmd_buffer->state.renderpass_cache);
cmd_buffer->usage_flags = pBeginInfo->flags;
tu_cs_begin(&cmd_buffer->cs);
tu_cs_begin(&cmd_buffer->draw_cs);
tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
- cmd_buffer->scratch_seqno = 0;
-
/* setup initial configuration into command buffer */
if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
switch (cmd_buffer->queue_family_index) {
default:
break;
}
- } else if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY &&
- (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT)) {
- assert(pBeginInfo->pInheritanceInfo);
- cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
- cmd_buffer->state.subpass = &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ } else if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
+ if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
+ assert(pBeginInfo->pInheritanceInfo);
+ cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
+ cmd_buffer->state.subpass =
+ &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ } else {
+ /* When executing in the middle of another command buffer, the CCU
+ * state is unknown.
+ */
+ cmd_buffer->state.ccu_state = TU_CMD_CCU_UNKNOWN;
+ }
}
cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
return VK_SUCCESS;
}
+/* Sets vertex buffers to HW binding points. We emit VBs in SDS (so that bin
+ * rendering can skip over unused state), so we need to collect all the
+ * bindings together into a single state emit at draw time.
+ */
void
tu_CmdBindVertexBuffers(VkCommandBuffer commandBuffer,
uint32_t firstBinding,
assert(firstBinding + bindingCount <= MAX_VBS);
for (uint32_t i = 0; i < bindingCount; i++) {
- cmd->state.vb.buffers[firstBinding + i] =
- tu_buffer_from_handle(pBuffers[i]);
+ struct tu_buffer *buf = tu_buffer_from_handle(pBuffers[i]);
+
+ cmd->state.vb.buffers[firstBinding + i] = buf;
cmd->state.vb.offsets[firstBinding + i] = pOffsets[i];
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
- /* VB states depend on VkPipelineVertexInputStateCreateInfo */
cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
}
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_buffer, buf, buffer);
- /* initialize/update the restart index */
- if (!cmd->state.index_buffer || cmd->state.index_type != indexType) {
- struct tu_cs *draw_cs = &cmd->draw_cs;
- tu6_emit_restart_index(
- draw_cs, indexType == VK_INDEX_TYPE_UINT32 ? 0xffffffff : 0xffff);
- tu_cs_sanity_check(draw_cs);
+ uint32_t index_size, index_shift, restart_index;
+
+ switch (indexType) {
+ case VK_INDEX_TYPE_UINT16:
+ index_size = INDEX4_SIZE_16_BIT;
+ index_shift = 1;
+ restart_index = 0xffff;
+ break;
+ case VK_INDEX_TYPE_UINT32:
+ index_size = INDEX4_SIZE_32_BIT;
+ index_shift = 2;
+ restart_index = 0xffffffff;
+ break;
+ case VK_INDEX_TYPE_UINT8_EXT:
+ index_size = INDEX4_SIZE_8_BIT;
+ index_shift = 0;
+ restart_index = 0xff;
+ break;
+ default:
+ unreachable("invalid VkIndexType");
}
- /* track the BO */
- if (cmd->state.index_buffer != buf)
- tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
+ /* initialize/update the restart index */
+ if (cmd->state.index_size != index_size)
+ tu_cs_emit_regs(&cmd->draw_cs, A6XX_PC_RESTART_INDEX(restart_index));
- cmd->state.index_buffer = buf;
- cmd->state.index_offset = offset;
- cmd->state.index_type = indexType;
+ assert(buf->size >= offset);
+
+ cmd->state.index_va = buf->bo->iova + buf->bo_offset + offset;
+ cmd->state.max_index_count = (buf->size - offset) >> index_shift;
+ cmd->state.index_size = index_size;
+ cmd->state.index_shift = index_shift;
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
void
uint32_t dynamicOffsetCount,
const uint32_t *pDynamicOffsets)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_pipeline_layout, layout, _layout);
unsigned dyn_idx = 0;
struct tu_descriptor_state *descriptors_state =
- tu_get_descriptors_state(cmd_buffer, pipelineBindPoint);
+ tu_get_descriptors_state(cmd, pipelineBindPoint);
for (unsigned i = 0; i < descriptorSetCount; ++i) {
unsigned idx = i + firstSet;
TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
descriptors_state->sets[idx] = set;
- descriptors_state->valid |= (1u << idx);
-
- /* Note: the actual input attachment indices come from the shader
- * itself, so we can't generate the patched versions of these until
- * draw time when both the pipeline and descriptors are bound and
- * we're inside the render pass.
- */
- unsigned dst_idx = layout->set[idx].input_attachment_start;
- memcpy(&descriptors_state->input_attachments[dst_idx * A6XX_TEX_CONST_DWORDS],
- set->dynamic_descriptors,
- set->layout->input_attachment_count * A6XX_TEX_CONST_DWORDS * 4);
for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
- /* Dynamic buffers come after input attachments in the descriptor set
- * itself, but due to how the Vulkan descriptor set binding works, we
- * have to put input attachments and dynamic buffers in separate
- * buffers in the descriptor_state and then combine them at draw
- * time. Binding a descriptor set only invalidates the descriptor
- * sets after it, but if we try to tightly pack the descriptors after
- * the input attachments then we could corrupt dynamic buffers in the
- * descriptor set before it, or we'd have to move all the dynamic
- * buffers over. We just put them into separate buffers to make
- * binding as well as the later patching of input attachments easy.
- */
- unsigned src_idx = j + set->layout->input_attachment_count;
+ /* update the contents of the dynamic descriptor set */
+ unsigned src_idx = j;
unsigned dst_idx = j + layout->set[idx].dynamic_offset_start;
assert(dyn_idx < dynamicOffsetCount);
dst[5] = va >> 32;
}
}
+
+ for (unsigned j = 0; j < set->layout->buffer_count; ++j) {
+ if (set->buffers[j]) {
+ tu_bo_list_add(&cmd->bo_list, set->buffers[j],
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+ }
+ }
+
+ if (set->size > 0) {
+ tu_bo_list_add(&cmd->bo_list, &set->pool->bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+ }
}
+ assert(dyn_idx == dynamicOffsetCount);
- if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE)
- cmd_buffer->state.dirty |= TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS;
- else
- cmd_buffer->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS;
+ uint32_t sp_bindless_base_reg, hlsq_bindless_base_reg, hlsq_update_value;
+ uint64_t addr[MAX_SETS + 1] = {};
+ struct tu_cs cs;
+
+ for (uint32_t i = 0; i < MAX_SETS; i++) {
+ struct tu_descriptor_set *set = descriptors_state->sets[i];
+ if (set)
+ addr[i] = set->va | 3;
+ }
+
+ if (layout->dynamic_offset_count) {
+ /* allocate and fill out dynamic descriptor set */
+ struct tu_cs_memory dynamic_desc_set;
+ VkResult result = tu_cs_alloc(&cmd->sub_cs, layout->dynamic_offset_count,
+ A6XX_TEX_CONST_DWORDS, &dynamic_desc_set);
+ assert(result == VK_SUCCESS);
+
+ memcpy(dynamic_desc_set.map, descriptors_state->dynamic_descriptors,
+ layout->dynamic_offset_count * A6XX_TEX_CONST_DWORDS * 4);
+ addr[MAX_SETS] = dynamic_desc_set.iova | 3;
+ }
+
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
+ sp_bindless_base_reg = REG_A6XX_SP_BINDLESS_BASE(0);
+ hlsq_bindless_base_reg = REG_A6XX_HLSQ_BINDLESS_BASE(0);
+ hlsq_update_value = 0x7c000;
+
+ cmd->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS | TU_CMD_DIRTY_SHADER_CONSTS;
+ } else {
+ assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE);
+
+ sp_bindless_base_reg = REG_A6XX_SP_CS_BINDLESS_BASE(0);
+ hlsq_bindless_base_reg = REG_A6XX_HLSQ_CS_BINDLESS_BASE(0);
+ hlsq_update_value = 0x3e00;
+
+ cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS;
+ }
+
+ tu_cs_begin_sub_stream(&cmd->sub_cs, 24, &cs);
+
+ tu_cs_emit_pkt4(&cs, sp_bindless_base_reg, 10);
+ tu_cs_emit_array(&cs, (const uint32_t*) addr, 10);
+ tu_cs_emit_pkt4(&cs, hlsq_bindless_base_reg, 10);
+ tu_cs_emit_array(&cs, (const uint32_t*) addr, 10);
+ tu_cs_emit_regs(&cs, A6XX_HLSQ_UPDATE_CNTL(.dword = hlsq_update_value));
+
+ struct tu_cs_entry ib = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
+ tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
+ tu_cs_emit_sds_ib(&cmd->draw_cs, TU_DRAW_STATE_DESC_SETS, ib);
+ cmd->state.desc_sets_ib = ib;
+ } else {
+ /* note: for compute we could emit directly, instead of a CP_INDIRECT
+ * however, the blob uses draw states for compute
+ */
+ tu_cs_emit_ib(&cmd->cs, &ib);
+ }
}
void tu_CmdBindTransformFeedbackBuffersEXT(VkCommandBuffer commandBuffer,
const VkDeviceSize *pSizes)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- assert(firstBinding + bindingCount <= IR3_MAX_SO_BUFFERS);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ /* using COND_REG_EXEC for xfb commands matches the blob behavior
+ * presumably there isn't any benefit using a draw state when the
+ * condition is (SYSMEM | BINNING)
+ */
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
for (uint32_t i = 0; i < bindingCount; i++) {
- uint32_t idx = firstBinding + i;
TU_FROM_HANDLE(tu_buffer, buf, pBuffers[i]);
+ uint64_t iova = buf->bo->iova + pOffsets[i];
+ uint32_t size = buf->bo->size - pOffsets[i];
+ uint32_t idx = i + firstBinding;
+
+ if (pSizes && pSizes[i] != VK_WHOLE_SIZE)
+ size = pSizes[i];
+
+ /* BUFFER_BASE is 32-byte aligned, add remaining offset to BUFFER_OFFSET */
+ uint32_t offset = iova & 0x1f;
+ iova &= ~(uint64_t) 0x1f;
- if (pOffsets[i] != 0)
- cmd->state.streamout_reset |= 1 << idx;
+ tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE(idx), 3);
+ tu_cs_emit_qw(cs, iova);
+ tu_cs_emit(cs, size + offset);
- cmd->state.streamout_buf.buffers[idx] = buf;
- cmd->state.streamout_buf.offsets[idx] = pOffsets[i];
- cmd->state.streamout_buf.sizes[idx] = pSizes[i];
+ cmd->state.streamout_offset[idx] = offset;
- cmd->state.streamout_enabled |= 1 << idx;
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
}
- cmd->state.dirty |= TU_CMD_DIRTY_STREAMOUT_BUFFERS;
+ tu_cond_exec_end(cs);
}
-void tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
- uint32_t firstCounterBuffer,
- uint32_t counterBufferCount,
- const VkBuffer *pCounterBuffers,
- const VkDeviceSize *pCounterBufferOffsets)
+void
+tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
{
- assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
- /* TODO do something with counter buffer? */
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
+
+ /* TODO: only update offset for active buffers */
+ for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++)
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, cmd->state.streamout_offset[i]));
+
+ for (uint32_t i = 0; i < counterBufferCount; i++) {
+ uint32_t idx = firstCounterBuffer + i;
+ uint32_t offset = cmd->state.streamout_offset[idx];
+
+ if (!pCounterBuffers[i])
+ continue;
+
+ TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
+
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+ tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+ CP_MEM_TO_REG_0_UNK31 |
+ CP_MEM_TO_REG_0_CNT(1));
+ tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+
+ if (offset) {
+ tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+ tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+ CP_REG_RMW_0_SRC1_ADD);
+ tu_cs_emit_qw(cs, 0xffffffff);
+ tu_cs_emit_qw(cs, offset);
+ }
+ }
+
+ tu_cond_exec_end(cs);
}
void tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
const VkBuffer *pCounterBuffers,
const VkDeviceSize *pCounterBufferOffsets)
{
- assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
- /* TODO do something with counter buffer? */
-
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- cmd->state.streamout_enabled = 0;
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
+
+ /* TODO: only flush buffers that need to be flushed */
+ for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
+ /* note: FLUSH_BASE is always the same, so it could go in init_hw()? */
+ tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_FLUSH_BASE(i), 2);
+ tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(flush_base[i]));
+ tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i);
+ }
+
+ for (uint32_t i = 0; i < counterBufferCount; i++) {
+ uint32_t idx = firstCounterBuffer + i;
+ uint32_t offset = cmd->state.streamout_offset[idx];
+
+ if (!pCounterBuffers[i])
+ continue;
+
+ TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
+
+ /* VPC_SO_FLUSH_BASE has dwords counter, but counter should be in bytes */
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+ tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_MEM_TO_REG_0_SHIFT_BY_2 |
+ 0x40000 | /* ??? */
+ CP_MEM_TO_REG_0_UNK31 |
+ CP_MEM_TO_REG_0_CNT(1));
+ tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(flush_base[idx]));
+
+ if (offset) {
+ tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+ tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_REG_RMW_0_SRC1_ADD);
+ tu_cs_emit_qw(cs, 0xffffffff);
+ tu_cs_emit_qw(cs, -offset);
+ }
+
+ tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
+ tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_REG_TO_MEM_0_CNT(1));
+ tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+ }
+
+ tu_cond_exec_end(cs);
+
+ cmd->state.xfb_used = true;
}
void
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
memcpy((void*) cmd->push_constants + offset, pValues, size);
- cmd->state.dirty |= TU_CMD_DIRTY_PUSH_CONSTANTS;
+ cmd->state.dirty |= TU_CMD_DIRTY_SHADER_CONSTS;
+}
+
+/* Flush everything which has been made available but we haven't actually
+ * flushed yet.
+ */
+static void
+tu_flush_all_pending(struct tu_cache_state *cache)
+{
+ cache->flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
+ cache->pending_flush_bits &= ~TU_CMD_FLAG_ALL_FLUSH;
}
VkResult
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- if (cmd_buffer->scratch_seqno) {
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->scratch_bo,
- MSM_SUBMIT_BO_WRITE);
+ /* We currently flush CCU at the end of the command buffer, like
+ * what the blob does. There's implicit synchronization around every
+ * vkQueueSubmit, but the kernel only flushes the UCHE, and we don't
+ * know yet if this command buffer will be the last in the submit so we
+ * have to defensively flush everything else.
+ *
+ * TODO: We could definitely do better than this, since these flushes
+ * aren't required by Vulkan, but we'd need kernel support to do that.
+ * Ideally, we'd like the kernel to flush everything afterwards, so that we
+ * wouldn't have to do any flushes here, and when submitting multiple
+ * command buffers there wouldn't be any unnecessary flushes in between.
+ */
+ if (cmd_buffer->state.pass) {
+ tu_flush_all_pending(&cmd_buffer->state.renderpass_cache);
+ tu_emit_cache_flush_renderpass(cmd_buffer, &cmd_buffer->draw_cs);
+ } else {
+ tu_flush_all_pending(&cmd_buffer->state.cache);
+ cmd_buffer->state.cache.flush_bits |=
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH;
+ tu_emit_cache_flush(cmd_buffer, &cmd_buffer->cs);
}
+ tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->scratch_bo,
+ MSM_SUBMIT_BO_WRITE);
+
if (cmd_buffer->use_vsc_data) {
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data,
+ tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_draw_strm,
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data2,
+ tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_prim_strm,
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
}
return cmd_buffer->record_result;
}
+static struct tu_cs
+tu_cmd_dynamic_state(struct tu_cmd_buffer *cmd, uint32_t id, uint32_t size)
+{
+ struct tu_cs_memory memory;
+ struct tu_cs cs;
+
+ /* TODO: share this logic with tu_pipeline_static_state */
+ tu_cs_alloc(&cmd->sub_cs, size, 1, &memory);
+ tu_cs_init_external(&cs, memory.map, memory.map + size);
+ tu_cs_begin(&cs);
+ tu_cs_reserve_space(&cs, size);
+
+ assert(id < ARRAY_SIZE(cmd->state.dynamic_state));
+ cmd->state.dynamic_state[id].iova = memory.iova;
+ cmd->state.dynamic_state[id].size = size;
+
+ tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
+ tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DYNAMIC + id, cmd->state.dynamic_state[id]);
+
+ return cs;
+}
+
void
tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
VkPipelineBindPoint pipelineBindPoint,
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_pipeline, pipeline, _pipeline);
- switch (pipelineBindPoint) {
- case VK_PIPELINE_BIND_POINT_GRAPHICS:
- cmd->state.pipeline = pipeline;
- cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
- break;
- case VK_PIPELINE_BIND_POINT_COMPUTE:
+ for (uint32_t i = 0; i < pipeline->cs.bo_count; i++) {
+ tu_bo_list_add(&cmd->bo_list, pipeline->cs.bos[i],
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+ }
+
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE) {
cmd->state.compute_pipeline = pipeline;
cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_PIPELINE;
- break;
- default:
- unreachable("unrecognized pipeline bind point");
- break;
+ return;
}
- tu_bo_list_add(&cmd->bo_list, &pipeline->program.binary_bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
- for (uint32_t i = 0; i < pipeline->cs.bo_count; i++) {
- tu_bo_list_add(&cmd->bo_list, pipeline->cs.bos[i],
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+ assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS);
+
+ cmd->state.pipeline = pipeline;
+ cmd->state.dirty |= TU_CMD_DIRTY_SHADER_CONSTS;
+
+ struct tu_cs *cs = &cmd->draw_cs;
+ uint32_t mask = ~pipeline->dynamic_state_mask & BITFIELD_MASK(TU_DYNAMIC_STATE_COUNT);
+ uint32_t i;
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (7 + util_bitcount(mask)));
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VI, pipeline->vi.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VI_BINNING, pipeline->vi.binning_state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_RAST, pipeline->rast.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DS, pipeline->ds.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_BLEND, pipeline->blend.state_ib);
+
+ for_each_bit(i, mask)
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i, pipeline->dynamic_state[i]);
+
+ /* If the new pipeline requires more VBs than we had previously set up, we
+ * need to re-emit them in SDS. If it requires the same set or fewer, we
+ * can just re-use the old SDS.
+ */
+ if (pipeline->vi.bindings_used & ~cmd->vertex_bindings_set)
+ cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
+
+ /* If the pipeline needs a dynamic descriptor, re-emit descriptor sets */
+ if (pipeline->layout->dynamic_offset_count)
+ cmd->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS;
+
+ /* dynamic linewidth state depends pipeline state's gras_su_cntl
+ * so the dynamic state ib must be updated when pipeline changes
+ */
+ if (pipeline->dynamic_state_mask & BIT(VK_DYNAMIC_STATE_LINE_WIDTH)) {
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_LINE_WIDTH, 2);
+
+ cmd->state.dynamic_gras_su_cntl &= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
+ cmd->state.dynamic_gras_su_cntl |= pipeline->gras_su_cntl;
+
+ tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = cmd->state.dynamic_gras_su_cntl));
}
}
const VkViewport *pViewports)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_VIEWPORT, 18);
assert(firstViewport == 0 && viewportCount == 1);
- cmd->state.dynamic.viewport.viewports[0] = pViewports[0];
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_VIEWPORT;
+
+ tu6_emit_viewport(&cs, pViewports);
}
void
const VkRect2D *pScissors)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_SCISSOR, 3);
assert(firstScissor == 0 && scissorCount == 1);
- cmd->state.dynamic.scissor.scissors[0] = pScissors[0];
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_SCISSOR;
+
+ tu6_emit_scissor(&cs, pScissors);
}
void
tu_CmdSetLineWidth(VkCommandBuffer commandBuffer, float lineWidth)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_LINE_WIDTH, 2);
- cmd->state.dynamic.line_width = lineWidth;
+ cmd->state.dynamic_gras_su_cntl &= ~A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
+ cmd->state.dynamic_gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(lineWidth / 2.0f);
- /* line width depends on VkPipelineRasterizationStateCreateInfo */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
+ tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = cmd->state.dynamic_gras_su_cntl));
}
void
float depthBiasSlopeFactor)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
-
- tu6_emit_depth_bias(draw_cs, depthBiasConstantFactor, depthBiasClamp,
- depthBiasSlopeFactor);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BIAS, 4);
- tu_cs_sanity_check(draw_cs);
+ tu6_emit_depth_bias(&cs, depthBiasConstantFactor, depthBiasClamp, depthBiasSlopeFactor);
}
void
const float blendConstants[4])
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_BLEND_CONSTANTS, 5);
- tu6_emit_blend_constants(draw_cs, blendConstants);
-
- tu_cs_sanity_check(draw_cs);
+ tu_cs_emit_pkt4(&cs, REG_A6XX_RB_BLEND_RED_F32, 4);
+ tu_cs_emit_array(&cs, (const uint32_t *) blendConstants, 4);
}
void
float minDepthBounds,
float maxDepthBounds)
{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BOUNDS, 3);
+
+ tu_cs_emit_regs(&cs,
+ A6XX_RB_Z_BOUNDS_MIN(minDepthBounds),
+ A6XX_RB_Z_BOUNDS_MAX(maxDepthBounds));
+}
+
+static void
+update_stencil_mask(uint32_t *value, VkStencilFaceFlags face, uint32_t mask)
+{
+ if (face & VK_STENCIL_FACE_FRONT_BIT)
+ *value |= A6XX_RB_STENCILMASK_MASK(mask);
+ if (face & VK_STENCIL_FACE_BACK_BIT)
+ *value |= A6XX_RB_STENCILMASK_BFMASK(mask);
}
void
uint32_t compareMask)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK, 2);
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_compare_mask.front = compareMask;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_compare_mask.back = compareMask;
+ update_stencil_mask(&cmd->state.dynamic_stencil_mask, faceMask, compareMask);
- /* the front/back compare masks must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILMASK(.dword = cmd->state.dynamic_stencil_mask));
}
void
uint32_t writeMask)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, 2);
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_write_mask.front = writeMask;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_write_mask.back = writeMask;
+ update_stencil_mask(&cmd->state.dynamic_stencil_wrmask, faceMask, writeMask);
- /* the front/back write masks must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILWRMASK(.dword = cmd->state.dynamic_stencil_wrmask));
}
void
uint32_t reference)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_REFERENCE, 2);
+
+ update_stencil_mask(&cmd->state.dynamic_stencil_ref, faceMask, reference);
+
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILREF(.dword = cmd->state.dynamic_stencil_ref));
+}
+
+void
+tu_CmdSetSampleLocationsEXT(VkCommandBuffer commandBuffer,
+ const VkSampleLocationsInfoEXT* pSampleLocationsInfo)
+{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_SAMPLE_LOCATIONS, 9);
+
+ assert(pSampleLocationsInfo);
+
+ tu6_emit_sample_locations(&cs, pSampleLocationsInfo);
+}
+
+static void
+tu_flush_for_access(struct tu_cache_state *cache,
+ enum tu_cmd_access_mask src_mask,
+ enum tu_cmd_access_mask dst_mask)
+{
+ enum tu_cmd_flush_bits flush_bits = 0;
+
+ if (src_mask & TU_ACCESS_SYSMEM_WRITE) {
+ cache->pending_flush_bits |= TU_CMD_FLAG_ALL_INVALIDATE;
+ }
+
+#define SRC_FLUSH(domain, flush, invalidate) \
+ if (src_mask & TU_ACCESS_##domain##_WRITE) { \
+ cache->pending_flush_bits |= TU_CMD_FLAG_##flush | \
+ (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate); \
+ }
+
+ SRC_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
+ SRC_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ SRC_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef SRC_FLUSH
+
+#define SRC_INCOHERENT_FLUSH(domain, flush, invalidate) \
+ if (src_mask & TU_ACCESS_##domain##_INCOHERENT_WRITE) { \
+ flush_bits |= TU_CMD_FLAG_##flush; \
+ cache->pending_flush_bits |= \
+ (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate); \
+ }
+
+ SRC_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ SRC_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef SRC_INCOHERENT_FLUSH
+
+ if (dst_mask & (TU_ACCESS_SYSMEM_READ | TU_ACCESS_SYSMEM_WRITE)) {
+ flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
+ }
+
+#define DST_FLUSH(domain, flush, invalidate) \
+ if (dst_mask & (TU_ACCESS_##domain##_READ | \
+ TU_ACCESS_##domain##_WRITE)) { \
+ flush_bits |= cache->pending_flush_bits & \
+ (TU_CMD_FLAG_##invalidate | \
+ (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush)); \
+ }
+
+ DST_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
+ DST_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ DST_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef DST_FLUSH
+
+#define DST_INCOHERENT_FLUSH(domain, flush, invalidate) \
+ if (dst_mask & (TU_ACCESS_##domain##_READ | \
+ TU_ACCESS_##domain##_WRITE)) { \
+ flush_bits |= TU_CMD_FLAG_##invalidate | \
+ (cache->pending_flush_bits & \
+ (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush)); \
+ }
+
+ DST_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ DST_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef DST_INCOHERENT_FLUSH
+
+ if (dst_mask & TU_ACCESS_WFI_READ) {
+ flush_bits |= TU_CMD_FLAG_WFI;
+ }
+
+ cache->flush_bits |= flush_bits;
+ cache->pending_flush_bits &= ~flush_bits;
+}
+
+static enum tu_cmd_access_mask
+vk2tu_access(VkAccessFlags flags, bool gmem)
+{
+ enum tu_cmd_access_mask mask = 0;
+
+ /* If the GPU writes a buffer that is then read by an indirect draw
+ * command, we theoretically need a WFI + WAIT_FOR_ME combination to
+ * wait for the writes to complete. The WAIT_FOR_ME is performed as part
+ * of the draw by the firmware, so we just need to execute a WFI.
+ */
+ if (flags &
+ (VK_ACCESS_INDIRECT_COMMAND_READ_BIT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_WFI_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_INDIRECT_COMMAND_READ_BIT | /* Read performed by CP */
+ VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT | /* Read performed by CP, I think */
+ VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT | /* Read performed by CP */
+ VK_ACCESS_HOST_READ_BIT | /* sysmem by definition */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_SYSMEM_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_HOST_WRITE_BIT |
+ VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_WRITE_BIT_EXT | /* Write performed by CP, I think */
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ }
+
+ if (flags &
+ (VK_ACCESS_INDEX_READ_BIT | /* Read performed by PC, I think */
+ VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | /* Read performed by VFD */
+ VK_ACCESS_UNIFORM_READ_BIT | /* Read performed by SP */
+ /* TODO: Is there a no-cache bit for textures so that we can ignore
+ * these?
+ */
+ VK_ACCESS_INPUT_ATTACHMENT_READ_BIT | /* Read performed by TP */
+ VK_ACCESS_SHADER_READ_BIT | /* Read perfomed by SP/TP */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_UCHE_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_SHADER_WRITE_BIT | /* Write performed by SP */
+ VK_ACCESS_TRANSFORM_FEEDBACK_WRITE_BIT_EXT | /* Write performed by VPC */
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_UCHE_WRITE;
+ }
+
+ /* When using GMEM, the CCU is always flushed automatically to GMEM, and
+ * then GMEM is flushed to sysmem. Furthermore, we already had to flush any
+ * previous writes in sysmem mode when transitioning to GMEM. Therefore we
+ * can ignore CCU and pretend that color attachments and transfers use
+ * sysmem directly.
+ */
+
+ if (flags &
+ (VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+ VK_ACCESS_COLOR_ATTACHMENT_READ_NONCOHERENT_BIT_EXT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ if (gmem)
+ mask |= TU_ACCESS_SYSMEM_READ;
+ else
+ mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ if (gmem)
+ mask |= TU_ACCESS_SYSMEM_READ;
+ else
+ mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ }
+ }
+
+ if (flags &
+ (VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
+ }
+ }
+
+ /* When the dst access is a transfer read/write, it seems we sometimes need
+ * to insert a WFI after any flushes, to guarantee that the flushes finish
+ * before the 2D engine starts. However the opposite (i.e. a WFI after
+ * CP_BLIT and before any subsequent flush) does not seem to be needed, and
+ * the blob doesn't emit such a WFI.
+ */
+
+ if (flags &
+ (VK_ACCESS_TRANSFER_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_COLOR_WRITE;
+ }
+ mask |= TU_ACCESS_WFI_READ;
+ }
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_reference.front = reference;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_reference.back = reference;
+ if (flags &
+ (VK_ACCESS_TRANSFER_READ_BIT | /* Access performed by TP */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_UCHE_READ | TU_ACCESS_WFI_READ;
+ }
- /* the front/back references must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
+ return mask;
}
+
void
tu_CmdExecuteCommands(VkCommandBuffer commandBuffer,
uint32_t commandBufferCount,
assert(commandBufferCount > 0);
+ /* Emit any pending flushes. */
+ if (cmd->state.pass) {
+ tu_flush_all_pending(&cmd->state.renderpass_cache);
+ tu_emit_cache_flush_renderpass(cmd, &cmd->draw_cs);
+ } else {
+ tu_flush_all_pending(&cmd->state.cache);
+ tu_emit_cache_flush(cmd, &cmd->cs);
+ }
+
for (uint32_t i = 0; i < commandBufferCount; i++) {
TU_FROM_HANDLE(tu_cmd_buffer, secondary, pCmdBuffers[i]);
cmd->record_result = result;
break;
}
+
+ if (secondary->has_tess)
+ cmd->has_tess = true;
} else {
assert(tu_cs_is_empty(&secondary->draw_cs));
assert(tu_cs_is_empty(&secondary->draw_epilogue_cs));
tu_cs_add_entries(&cmd->cs, &secondary->cs);
}
+
+ cmd->state.index_size = secondary->state.index_size; /* for restart index update */
}
cmd->state.dirty = ~0u; /* TODO: set dirty only what needs to be */
+
+ /* After executing secondary command buffers, there may have been arbitrary
+ * flushes executed, so when we encounter a pipeline barrier with a
+ * srcMask, we have to assume that we need to invalidate. Therefore we need
+ * to re-initialize the cache with all pending invalidate bits set.
+ */
+ if (cmd->state.pass) {
+ tu_cache_init(&cmd->state.renderpass_cache);
+ } else {
+ tu_cache_init(&cmd->state.cache);
+ }
}
VkResult
}
}
+static void
+tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
+ const struct tu_subpass_barrier *barrier,
+ bool external)
+{
+ /* Note: we don't know until the end of the subpass whether we'll use
+ * sysmem, so assume sysmem here to be safe.
+ */
+ struct tu_cache_state *cache =
+ external ? &cmd_buffer->state.cache : &cmd_buffer->state.renderpass_cache;
+ enum tu_cmd_access_mask src_flags =
+ vk2tu_access(barrier->src_access_mask, false);
+ enum tu_cmd_access_mask dst_flags =
+ vk2tu_access(barrier->dst_access_mask, false);
+
+ if (barrier->incoherent_ccu_color)
+ src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ if (barrier->incoherent_ccu_depth)
+ src_flags |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
+
+ tu_flush_for_access(cache, src_flags, dst_flags);
+}
+
void
tu_CmdBeginRenderPass(VkCommandBuffer commandBuffer,
const VkRenderPassBeginInfo *pRenderPassBegin,
tu_cmd_update_tiling_config(cmd, &pRenderPassBegin->renderArea);
tu_cmd_prepare_tile_store_ib(cmd);
- tu_emit_load_clear(cmd, pRenderPassBegin);
+ /* Note: because this is external, any flushes will happen before draw_cs
+ * gets called. However deferred flushes could have to happen later as part
+ * of the subpass.
+ */
+ tu_subpass_barrier(cmd, &pass->subpasses[0].start_barrier, true);
+ cmd->state.renderpass_cache.pending_flush_bits =
+ cmd->state.cache.pending_flush_bits;
+ cmd->state.renderpass_cache.flush_bits = 0;
+
+ tu_emit_renderpass_begin(cmd, pRenderPassBegin);
tu6_emit_zs(cmd, cmd->state.subpass, &cmd->draw_cs);
tu6_emit_mrt(cmd, cmd->state.subpass, &cmd->draw_cs);
tu6_emit_msaa(&cmd->draw_cs, cmd->state.subpass->samples);
tu6_emit_render_cntl(cmd, cmd->state.subpass, &cmd->draw_cs, false);
- /* note: use_hw_binning only checks tiling config */
- if (use_hw_binning(cmd))
- cmd->use_vsc_data = true;
+ tu_set_input_attachments(cmd, cmd->state.subpass);
for (uint32_t i = 0; i < fb->attachment_count; ++i) {
const struct tu_image_view *iview = fb->attachments[i].attachment;
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
}
- /* Flag input attachment descriptors for re-emission if necessary */
- cmd->state.dirty |= TU_CMD_DIRTY_INPUT_ATTACHMENTS;
+ cmd->state.dirty |= TU_CMD_DIRTY_DRAW_STATE;
}
void
tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
- /* Emit flushes so that input attachments will read the correct value.
- * TODO: use subpass dependencies to flush or not
- */
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
-
- if (subpass->resolve_attachments) {
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
-
- for (unsigned i = 0; i < subpass->color_count; i++) {
- uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a == VK_ATTACHMENT_UNUSED)
- continue;
-
- tu6_emit_sysmem_resolve(cmd, cs, a,
- subpass->color_attachments[i].attachment);
- }
-
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- }
+ tu6_emit_sysmem_resolves(cmd, cs, subpass);
tu_cond_exec_end(cs);
- /* subpass->input_count > 0 then texture cache invalidate is likely to be needed */
- if (cmd->state.subpass->input_count)
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
+ /* Handle dependencies for the next subpass */
+ tu_subpass_barrier(cmd, &cmd->state.subpass->start_barrier, false);
/* emit mrt/zs/msaa/ubwc state for the subpass that is starting */
tu6_emit_zs(cmd, cmd->state.subpass, cs);
tu6_emit_msaa(cs, cmd->state.subpass->samples);
tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, false);
- /* Flag input attachment descriptors for re-emission if necessary */
- cmd->state.dirty |= TU_CMD_DIRTY_INPUT_ATTACHMENTS;
+ tu_set_input_attachments(cmd, cmd->state.subpass);
}
void
tu_CmdNextSubpass(commandBuffer, pSubpassBeginInfo->contents);
}
-struct tu_draw_info
+static void
+tu6_emit_user_consts(struct tu_cs *cs, const struct tu_pipeline *pipeline,
+ struct tu_descriptor_state *descriptors_state,
+ gl_shader_stage type,
+ uint32_t *push_constants)
{
- /**
- * Number of vertices.
- */
- uint32_t count;
-
- /**
- * Index of the first vertex.
- */
- int32_t vertex_offset;
-
- /**
- * First instance id.
- */
- uint32_t first_instance;
-
- /**
- * Number of instances.
- */
- uint32_t instance_count;
-
- /**
- * First index (indexed draws only).
- */
- uint32_t first_index;
-
- /**
- * Whether it's an indexed draw.
- */
- bool indexed;
-
- /**
- * Indirect draw parameters resource.
- */
- struct tu_buffer *indirect;
- uint64_t indirect_offset;
- uint32_t stride;
-
- /**
- * Draw count parameters resource.
- */
- struct tu_buffer *count_buffer;
- uint64_t count_buffer_offset;
-
- /**
- * Stream output parameters resource.
- */
- struct tu_buffer *streamout_buffer;
- uint64_t streamout_buffer_offset;
-};
-
-#define ENABLE_ALL (CP_SET_DRAW_STATE__0_BINNING | CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
-#define ENABLE_DRAW (CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
-#define ENABLE_NON_GMEM (CP_SET_DRAW_STATE__0_BINNING | CP_SET_DRAW_STATE__0_SYSMEM)
-
-enum tu_draw_state_group_id
-{
- TU_DRAW_STATE_PROGRAM,
- TU_DRAW_STATE_PROGRAM_BINNING,
- TU_DRAW_STATE_VI,
- TU_DRAW_STATE_VI_BINNING,
- TU_DRAW_STATE_VP,
- TU_DRAW_STATE_RAST,
- TU_DRAW_STATE_DS,
- TU_DRAW_STATE_BLEND,
- TU_DRAW_STATE_VS_CONST,
- TU_DRAW_STATE_GS_CONST,
- TU_DRAW_STATE_FS_CONST,
- TU_DRAW_STATE_DESC_SETS,
- TU_DRAW_STATE_DESC_SETS_GMEM,
- TU_DRAW_STATE_DESC_SETS_LOAD,
- TU_DRAW_STATE_VS_PARAMS,
-
- TU_DRAW_STATE_COUNT,
-};
-
-struct tu_draw_state_group
-{
- enum tu_draw_state_group_id id;
- uint32_t enable_mask;
- struct tu_cs_entry ib;
-};
-
-static inline uint32_t
-tu6_stage2opcode(gl_shader_stage type)
-{
- switch (type) {
- case MESA_SHADER_VERTEX:
- case MESA_SHADER_TESS_CTRL:
- case MESA_SHADER_TESS_EVAL:
- case MESA_SHADER_GEOMETRY:
- return CP_LOAD_STATE6_GEOM;
- case MESA_SHADER_FRAGMENT:
- case MESA_SHADER_COMPUTE:
- case MESA_SHADER_KERNEL:
- return CP_LOAD_STATE6_FRAG;
- default:
- unreachable("bad shader type");
- }
-}
-
-static inline enum a6xx_state_block
-tu6_stage2shadersb(gl_shader_stage type)
-{
- switch (type) {
- case MESA_SHADER_VERTEX:
- return SB6_VS_SHADER;
- case MESA_SHADER_GEOMETRY:
- return SB6_GS_SHADER;
- case MESA_SHADER_FRAGMENT:
- return SB6_FS_SHADER;
- case MESA_SHADER_COMPUTE:
- case MESA_SHADER_KERNEL:
- return SB6_CS_SHADER;
- default:
- unreachable("bad shader type");
- return ~0;
- }
-}
-
-static void
-tu6_emit_user_consts(struct tu_cs *cs, const struct tu_pipeline *pipeline,
- struct tu_descriptor_state *descriptors_state,
- gl_shader_stage type,
- uint32_t *push_constants)
-{
- const struct tu_program_descriptor_linkage *link =
- &pipeline->program.link[type];
- const struct ir3_ubo_analysis_state *state = &link->ubo_state;
+ const struct tu_program_descriptor_linkage *link =
+ &pipeline->program.link[type];
+ const struct ir3_ubo_analysis_state *state = &link->const_state.ubo_state;
if (link->push_consts.count > 0) {
unsigned num_units = link->push_consts.count;
/* Dig out the descriptor from the descriptor state and read the VA from
* it.
*/
- assert(state->range[i].bindless);
- uint32_t *base = state->range[i].bindless_base == MAX_SETS ?
+ assert(state->range[i].ubo.bindless);
+ uint32_t *base = state->range[i].ubo.bindless_base == MAX_SETS ?
descriptors_state->dynamic_descriptors :
- descriptors_state->sets[state->range[i].bindless_base]->mapped_ptr;
- unsigned block = state->range[i].block;
- /* If the block in the shader here is in the dynamic descriptor set, it
- * is an index into the dynamic descriptor set which is combined from
- * dynamic descriptors and input attachments on-the-fly, and we don't
- * have access to it here. Instead we work backwards to get the index
- * into dynamic_descriptors.
- */
- if (state->range[i].bindless_base == MAX_SETS)
- block -= pipeline->layout->input_attachment_count;
+ descriptors_state->sets[state->range[i].ubo.bindless_base]->mapped_ptr;
+ unsigned block = state->range[i].ubo.block;
uint32_t *desc = base + block * A6XX_TEX_CONST_DWORDS;
uint64_t va = desc[0] | ((uint64_t)(desc[1] & A6XX_UBO_1_BASE_HI__MASK) << 32);
assert(va);
static VkResult
tu6_emit_vs_params(struct tu_cmd_buffer *cmd,
- const struct tu_draw_info *draw,
+ uint32_t first_instance,
struct tu_cs_entry *entry)
{
/* TODO: fill out more than just base instance */
tu_cs_emit(&cs, 0);
tu_cs_emit(&cs, 0);
- tu_cs_emit(&cs, draw->first_instance);
+ tu_cs_emit(&cs, first_instance);
tu_cs_emit(&cs, 0);
*entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
return VK_SUCCESS;
}
-static VkResult
-tu6_emit_descriptor_sets(struct tu_cmd_buffer *cmd,
- const struct tu_pipeline *pipeline,
- VkPipelineBindPoint bind_point,
- struct tu_cs_entry *entry,
- bool gmem)
-{
- struct tu_cs *draw_state = &cmd->sub_cs;
- struct tu_pipeline_layout *layout = pipeline->layout;
- struct tu_descriptor_state *descriptors_state =
- tu_get_descriptors_state(cmd, bind_point);
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
- const uint32_t *input_attachment_idx =
- pipeline->program.input_attachment_idx;
- uint32_t num_dynamic_descs = layout->dynamic_offset_count +
- layout->input_attachment_count;
- struct ts_cs_memory dynamic_desc_set;
- VkResult result;
+static struct tu_cs_entry
+tu6_emit_vertex_buffers(struct tu_cmd_buffer *cmd,
+ const struct tu_pipeline *pipeline)
+{
+ struct tu_cs cs;
+ tu_cs_begin_sub_stream(&cmd->sub_cs, 4 * MAX_VBS, &cs);
- if (num_dynamic_descs > 0) {
- /* allocate and fill out dynamic descriptor set */
- result = tu_cs_alloc(draw_state, num_dynamic_descs,
- A6XX_TEX_CONST_DWORDS, &dynamic_desc_set);
- if (result != VK_SUCCESS)
- return result;
+ int binding;
+ for_each_bit(binding, pipeline->vi.bindings_used) {
+ const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
+ const VkDeviceSize offset = buf->bo_offset +
+ cmd->state.vb.offsets[binding];
- memcpy(dynamic_desc_set.map, descriptors_state->input_attachments,
- layout->input_attachment_count * A6XX_TEX_CONST_DWORDS * 4);
+ tu_cs_emit_regs(&cs,
+ A6XX_VFD_FETCH_BASE(binding, .bo = buf->bo, .bo_offset = offset),
+ A6XX_VFD_FETCH_SIZE(binding, buf->size - offset));
- if (gmem) {
- /* Patch input attachments to refer to GMEM instead */
- for (unsigned i = 0; i < layout->input_attachment_count; i++) {
- uint32_t *dst =
- &dynamic_desc_set.map[A6XX_TEX_CONST_DWORDS * i];
-
- /* The compiler has already laid out input_attachment_idx in the
- * final order of input attachments, so there's no need to go
- * through the pipeline layout finding input attachments.
- */
- unsigned attachment_idx = input_attachment_idx[i];
+ }
- /* It's possible for the pipeline layout to include an input
- * attachment which doesn't actually exist for the current
- * subpass. Of course, this is only valid so long as the pipeline
- * doesn't try to actually load that attachment. Just skip
- * patching in that scenario to avoid out-of-bounds accesses.
- */
- if (attachment_idx >= cmd->state.subpass->input_count)
- continue;
-
- uint32_t a = cmd->state.subpass->input_attachments[attachment_idx].attachment;
- const struct tu_render_pass_attachment *att = &cmd->state.pass->attachments[a];
-
- assert(att->gmem_offset >= 0);
-
- dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
- dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
- dst[2] &= ~(A6XX_TEX_CONST_2_TYPE__MASK | A6XX_TEX_CONST_2_PITCH__MASK);
- dst[2] |=
- A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
- A6XX_TEX_CONST_2_PITCH(tiling->tile0.extent.width * att->cpp);
- dst[3] = 0;
- dst[4] = cmd->device->physical_device->gmem_base + att->gmem_offset;
- dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
- for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
- dst[i] = 0;
-
- if (cmd->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY)
- tu_finishme("patch input attachment pitch for secondary cmd buffer");
- }
- }
+ cmd->vertex_bindings_set = pipeline->vi.bindings_used;
- memcpy(dynamic_desc_set.map + layout->input_attachment_count * A6XX_TEX_CONST_DWORDS,
- descriptors_state->dynamic_descriptors,
- layout->dynamic_offset_count * A6XX_TEX_CONST_DWORDS * 4);
- }
+ return tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
+}
- uint32_t sp_bindless_base_reg, hlsq_bindless_base_reg;
- uint32_t hlsq_update_value;
- switch (bind_point) {
- case VK_PIPELINE_BIND_POINT_GRAPHICS:
- sp_bindless_base_reg = REG_A6XX_SP_BINDLESS_BASE(0);
- hlsq_bindless_base_reg = REG_A6XX_HLSQ_BINDLESS_BASE(0);
- hlsq_update_value = 0x7c000;
+static uint64_t
+get_tess_param_bo_size(const struct tu_pipeline *pipeline,
+ uint32_t draw_count)
+{
+ /* TODO: For indirect draws, we can't compute the BO size ahead of time.
+ * Still not sure what to do here, so just allocate a reasonably large
+ * BO and hope for the best for now.
+ * (maxTessellationControlPerVertexOutputComponents * 2048 vertices +
+ * maxTessellationControlPerPatchOutputComponents * 512 patches) */
+ if (!draw_count) {
+ return ((128 * 2048) + (128 * 512)) * 4;
+ }
+
+ /* For each patch, adreno lays out the tess param BO in memory as:
+ * (v_input[0][0])...(v_input[i][j])(p_input[0])...(p_input[k]).
+ * where i = # vertices per patch, j = # per-vertex outputs, and
+ * k = # per-patch outputs.*/
+ uint32_t verts_per_patch = pipeline->ia.primtype - DI_PT_PATCHES0;
+ uint32_t num_patches = draw_count / verts_per_patch;
+ return draw_count * pipeline->tess.per_vertex_output_size +
+ pipeline->tess.per_patch_output_size * num_patches;
+}
+
+static uint64_t
+get_tess_factor_bo_size(const struct tu_pipeline *pipeline,
+ uint32_t draw_count)
+{
+ /* TODO: For indirect draws, we can't compute the BO size ahead of time.
+ * Still not sure what to do here, so just allocate a reasonably large
+ * BO and hope for the best for now.
+ * (quad factor stride * 512 patches) */
+ if (!draw_count) {
+ return (28 * 512) * 4;
+ }
+
+ /* Each distinct patch gets its own tess factor output. */
+ uint32_t verts_per_patch = pipeline->ia.primtype - DI_PT_PATCHES0;
+ uint32_t num_patches = draw_count / verts_per_patch;
+ uint32_t factor_stride;
+ switch (pipeline->tess.patch_type) {
+ case IR3_TESS_ISOLINES:
+ factor_stride = 12;
break;
- case VK_PIPELINE_BIND_POINT_COMPUTE:
- sp_bindless_base_reg = REG_A6XX_SP_CS_BINDLESS_BASE(0);
- hlsq_bindless_base_reg = REG_A6XX_HLSQ_CS_BINDLESS_BASE(0);
- hlsq_update_value = 0x3e00;
+ case IR3_TESS_TRIANGLES:
+ factor_stride = 20;
+ break;
+ case IR3_TESS_QUADS:
+ factor_stride = 28;
break;
default:
- unreachable("bad bind point");
+ unreachable("bad tessmode");
}
+ return factor_stride * num_patches;
+}
- /* Be careful here to *not* refer to the pipeline, so that if only the
- * pipeline changes we don't have to emit this again (except if there are
- * dynamic descriptors in the pipeline layout). This means always emitting
- * all the valid descriptors, which means that we always have to put the
- * dynamic descriptor in the driver-only slot at the end
- */
- uint32_t num_user_sets = util_last_bit(descriptors_state->valid);
- uint32_t num_sets = num_user_sets;
- if (num_dynamic_descs > 0) {
- num_user_sets = MAX_SETS;
- num_sets = num_user_sets + 1;
- }
-
- unsigned regs[2] = { sp_bindless_base_reg, hlsq_bindless_base_reg };
-
+static VkResult
+tu6_emit_tess_consts(struct tu_cmd_buffer *cmd,
+ uint32_t draw_count,
+ const struct tu_pipeline *pipeline,
+ struct tu_cs_entry *entry)
+{
struct tu_cs cs;
- result = tu_cs_begin_sub_stream(draw_state, ARRAY_SIZE(regs) * (1 + num_sets * 2) + 2, &cs);
+ VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 20, &cs);
if (result != VK_SUCCESS)
return result;
- if (num_sets > 0) {
- for (unsigned i = 0; i < ARRAY_SIZE(regs); i++) {
- tu_cs_emit_pkt4(&cs, regs[i], num_sets * 2);
- for (unsigned j = 0; j < num_user_sets; j++) {
- if (descriptors_state->valid & (1 << j)) {
- /* magic | 3 copied from the blob */
- tu_cs_emit_qw(&cs, descriptors_state->sets[j]->va | 3);
- } else {
- tu_cs_emit_qw(&cs, 0 | 3);
- }
- }
- if (num_dynamic_descs > 0) {
- tu_cs_emit_qw(&cs, dynamic_desc_set.iova | 3);
- }
- }
-
- tu_cs_emit_regs(&cs, A6XX_HLSQ_UPDATE_CNTL(hlsq_update_value));
- }
-
- *entry = tu_cs_end_sub_stream(draw_state, &cs);
- return VK_SUCCESS;
-}
-
-static void
-tu6_emit_streamout(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- struct tu_streamout_state *tf = &cmd->state.pipeline->streamout;
-
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
- if (!buf)
- continue;
-
- uint32_t offset;
- offset = cmd->state.streamout_buf.offsets[i];
-
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_BASE(i, .bo = buf->bo,
- .bo_offset = buf->bo_offset));
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_SIZE(i, buf->size));
-
- if (cmd->state.streamout_reset & (1 << i)) {
- offset *= tf->stride[i];
+ uint64_t tess_factor_size = get_tess_factor_bo_size(pipeline, draw_count);
+ uint64_t tess_param_size = get_tess_param_bo_size(pipeline, draw_count);
+ uint64_t tess_bo_size = tess_factor_size + tess_param_size;
+ if (tess_bo_size > 0) {
+ struct tu_bo *tess_bo;
+ result = tu_get_scratch_bo(cmd->device, tess_bo_size, &tess_bo);
+ if (result != VK_SUCCESS)
+ return result;
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, offset));
- cmd->state.streamout_reset &= ~(1 << i);
- } else {
- tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
- tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(i)) |
- CP_MEM_TO_REG_0_SHIFT_BY_2 | CP_MEM_TO_REG_0_UNK31 |
- CP_MEM_TO_REG_0_CNT(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova +
- ctrl_offset(flush_base[i].offset));
- }
+ tu_bo_list_add(&cmd->bo_list, tess_bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+ uint64_t tess_factor_iova = tess_bo->iova;
+ uint64_t tess_param_iova = tess_factor_iova + tess_factor_size;
- tu_cs_emit_regs(cs, A6XX_VPC_SO_FLUSH_BASE(i, .bo = &cmd->scratch_bo,
- .bo_offset =
- ctrl_offset(flush_base[i])));
- }
-
- if (cmd->state.streamout_enabled) {
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 12 + (2 * tf->prog_count));
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, tf->vpc_so_buf_cntl);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(0));
- tu_cs_emit(cs, tf->ncomp[0]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(1));
- tu_cs_emit(cs, tf->ncomp[1]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(2));
- tu_cs_emit(cs, tf->ncomp[2]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(3));
- tu_cs_emit(cs, tf->ncomp[3]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, A6XX_VPC_SO_CNTL_ENABLE);
- for (unsigned i = 0; i < tf->prog_count; i++) {
- tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
- tu_cs_emit(cs, tf->prog[i]);
- }
- } else {
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, 0);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, 0);
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(pipeline->tess.hs_bo_regid) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_HS_SHADER) |
+ CP_LOAD_STATE6_0_NUM_UNIT(1));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+ tu_cs_emit_qw(&cs, tess_param_iova);
+ tu_cs_emit_qw(&cs, tess_factor_iova);
+
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(pipeline->tess.ds_bo_regid) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_DS_SHADER) |
+ CP_LOAD_STATE6_0_NUM_UNIT(1));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+ tu_cs_emit_qw(&cs, tess_param_iova);
+ tu_cs_emit_qw(&cs, tess_factor_iova);
+
+ tu_cs_emit_pkt4(&cs, REG_A6XX_PC_TESSFACTOR_ADDR_LO, 2);
+ tu_cs_emit_qw(&cs, tess_factor_iova);
+
+ /* TODO: Without this WFI here, the hardware seems unable to read these
+ * addresses we just emitted. Freedreno emits these consts as part of
+ * IB1 instead of in a draw state which might make this WFI unnecessary,
+ * but it requires a bit more indirection (SS6_INDIRECT for consts). */
+ tu_cs_emit_wfi(&cs);
}
+ *entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
+ return VK_SUCCESS;
}
static VkResult
-tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
+tu6_draw_common(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ bool indexed,
+ uint32_t vertex_offset,
+ uint32_t first_instance,
+ /* note: draw_count count is 0 for indirect */
+ uint32_t draw_count)
{
const struct tu_pipeline *pipeline = cmd->state.pipeline;
- const struct tu_dynamic_state *dynamic = &cmd->state.dynamic;
- struct tu_draw_state_group draw_state_groups[TU_DRAW_STATE_COUNT];
- uint32_t draw_state_group_count = 0;
VkResult result;
struct tu_descriptor_state *descriptors_state =
&cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
+ tu_emit_cache_flush_renderpass(cmd, cs);
+
/* TODO lrz */
tu_cs_emit_regs(cs,
- A6XX_PC_PRIMITIVE_CNTL_0(.primitive_restart =
- pipeline->ia.primitive_restart && draw->indexed));
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH)) {
- tu6_emit_gras_su_cntl(cs, pipeline->rast.gras_su_cntl,
- dynamic->line_width);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
- tu6_emit_stencil_compare_mask(cs, dynamic->stencil_compare_mask.front,
- dynamic->stencil_compare_mask.back);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
- tu6_emit_stencil_write_mask(cs, dynamic->stencil_write_mask.front,
- dynamic->stencil_write_mask.back);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
- tu6_emit_stencil_reference(cs, dynamic->stencil_reference.front,
- dynamic->stencil_reference.back);
+ A6XX_VFD_INDEX_OFFSET(vertex_offset),
+ A6XX_VFD_INSTANCE_START_OFFSET(first_instance));
+
+ tu_cs_emit_regs(cs, A6XX_PC_PRIMITIVE_CNTL_0(
+ .primitive_restart =
+ pipeline->ia.primitive_restart && indexed,
+ .tess_upper_left_domain_origin =
+ pipeline->tess.upper_left_domain_origin));
+
+ if (cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) {
+ cmd->state.shader_const_ib[MESA_SHADER_VERTEX] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_VERTEX);
+ cmd->state.shader_const_ib[MESA_SHADER_TESS_CTRL] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_TESS_CTRL);
+ cmd->state.shader_const_ib[MESA_SHADER_TESS_EVAL] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_TESS_EVAL);
+ cmd->state.shader_const_ib[MESA_SHADER_GEOMETRY] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_GEOMETRY);
+ cmd->state.shader_const_ib[MESA_SHADER_FRAGMENT] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT);
}
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_VIEWPORT) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_VIEWPORT)) {
- tu6_emit_viewport(cs, &cmd->state.dynamic.viewport.viewports[0]);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_SCISSOR) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_SCISSOR)) {
- tu6_emit_scissor(cs, &cmd->state.dynamic.scissor.scissors[0]);
- }
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_VERTEX_BUFFERS)) {
- for (uint32_t i = 0; i < pipeline->vi.count; i++) {
- const uint32_t binding = pipeline->vi.bindings[i];
- const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
- const VkDeviceSize offset = buf->bo_offset +
- cmd->state.vb.offsets[binding];
- const VkDeviceSize size =
- offset < buf->size ? buf->size - offset : 0;
-
- tu_cs_emit_regs(cs,
- A6XX_VFD_FETCH_BASE(i, .bo = buf->bo, .bo_offset = offset),
- A6XX_VFD_FETCH_SIZE(i, size));
- }
- }
-
- if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_PROGRAM,
- .enable_mask = ENABLE_DRAW,
- .ib = pipeline->program.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_PROGRAM_BINNING,
- .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
- .ib = pipeline->program.binning_state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VI,
- .enable_mask = ENABLE_DRAW,
- .ib = pipeline->vi.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VI_BINNING,
- .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
- .ib = pipeline->vi.binning_state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VP,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->vp.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_RAST,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->rast.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_DS,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->ds.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_BLEND,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->blend.state_ib,
- };
- }
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS | TU_CMD_DIRTY_PUSH_CONSTANTS)) {
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VS_CONST,
- .enable_mask = ENABLE_ALL,
- .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_VERTEX)
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_GS_CONST,
- .enable_mask = ENABLE_ALL,
- .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_GEOMETRY)
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_FS_CONST,
- .enable_mask = ENABLE_DRAW,
- .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT)
- };
- }
-
- if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS)
- tu6_emit_streamout(cmd, cs);
-
- /* If there are any any dynamic descriptors, then we may need to re-emit
- * them after every pipeline change in case the number of input attachments
- * changes. We also always need to re-emit after a pipeline change if there
- * are any input attachments, because the input attachment index comes from
- * the pipeline. Finally, it can also happen that the subpass changes
- * without the pipeline changing, in which case the GMEM descriptors need
- * to be patched differently.
- *
- * TODO: We could probably be clever and avoid re-emitting state on
- * pipeline changes if the number of input attachments is always 0. We
- * could also only re-emit dynamic state.
- */
- if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS ||
- ((pipeline->layout->dynamic_offset_count +
- pipeline->layout->input_attachment_count > 0) &&
- cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) ||
- (pipeline->layout->input_attachment_count > 0 &&
- cmd->state.dirty & TU_CMD_DIRTY_INPUT_ATTACHMENTS)) {
- struct tu_cs_entry desc_sets, desc_sets_gmem;
- bool need_gmem_desc_set = pipeline->layout->input_attachment_count > 0;
-
- result = tu6_emit_descriptor_sets(cmd, pipeline,
- VK_PIPELINE_BIND_POINT_GRAPHICS,
- &desc_sets, false);
- if (result != VK_SUCCESS)
- return result;
-
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_DESC_SETS,
- .enable_mask = need_gmem_desc_set ? ENABLE_NON_GMEM : ENABLE_ALL,
- .ib = desc_sets,
- };
-
- if (need_gmem_desc_set) {
- result = tu6_emit_descriptor_sets(cmd, pipeline,
- VK_PIPELINE_BIND_POINT_GRAPHICS,
- &desc_sets_gmem, true);
- if (result != VK_SUCCESS)
- return result;
-
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_DESC_SETS_GMEM,
- .enable_mask = CP_SET_DRAW_STATE__0_GMEM,
- .ib = desc_sets_gmem,
- };
- }
-
+ if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
/* We need to reload the descriptors every time the descriptor sets
* change. However, the commands we send only depend on the pipeline
* because the whole point is to cache descriptors which are used by the
tu_cs_emit_array(&load_cs,
(uint32_t *)((char *)load_entry->bo->map + load_entry->offset),
load_entry->size / 4);
- struct tu_cs_entry load_copy = tu_cs_end_sub_stream(&cmd->sub_cs, &load_cs);
-
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_DESC_SETS_LOAD,
- /* The blob seems to not enable this for binning, even when
- * resources would actually be used in the binning shader.
- * Presumably the overhead of prefetching the resources isn't
- * worth it.
- */
- .enable_mask = ENABLE_DRAW,
- .ib = load_copy,
- };
+ cmd->state.desc_sets_load_ib = tu_cs_end_sub_stream(&cmd->sub_cs, &load_cs);
+ } else {
+ cmd->state.desc_sets_load_ib.size = 0;
}
}
+ if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS)
+ cmd->state.vertex_buffers_ib = tu6_emit_vertex_buffers(cmd, pipeline);
+
struct tu_cs_entry vs_params;
- result = tu6_emit_vs_params(cmd, draw, &vs_params);
+ result = tu6_emit_vs_params(cmd, first_instance, &vs_params);
if (result != VK_SUCCESS)
return result;
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VS_PARAMS,
- .enable_mask = ENABLE_ALL,
- .ib = vs_params,
- };
+ bool has_tess =
+ pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
+ struct tu_cs_entry tess_consts = {};
+ if (has_tess) {
+ cmd->has_tess = true;
+ result = tu6_emit_tess_consts(cmd, draw_count, pipeline, &tess_consts);
+ if (result != VK_SUCCESS)
+ return result;
+ }
- tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_group_count);
- for (uint32_t i = 0; i < draw_state_group_count; i++) {
- const struct tu_draw_state_group *group = &draw_state_groups[i];
- debug_assert((group->enable_mask & ~ENABLE_ALL) == 0);
- uint32_t cp_set_draw_state =
- CP_SET_DRAW_STATE__0_COUNT(group->ib.size / 4) |
- group->enable_mask |
- CP_SET_DRAW_STATE__0_GROUP_ID(group->id);
- uint64_t iova;
- if (group->ib.size) {
- iova = group->ib.bo->iova + group->ib.offset;
- } else {
- cp_set_draw_state |= CP_SET_DRAW_STATE__0_DISABLE;
- iova = 0;
+ /* for the first draw in a renderpass, re-emit all the draw states
+ *
+ * and if a draw-state disabling path (CmdClearAttachments 3D fallback) was
+ * used, then draw states must be re-emitted. note however this only happens
+ * in the sysmem path, so this can be skipped this for the gmem path (TODO)
+ *
+ * the two input attachment states are excluded because secondary command
+ * buffer doesn't have a state ib to restore it, and not re-emitting them
+ * is OK since CmdClearAttachments won't disable/overwrite them
+ */
+ if (cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE) {
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (TU_DRAW_STATE_COUNT - 2));
+
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_TESS, tess_consts);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VI, pipeline->vi.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VI_BINNING, pipeline->vi.binning_state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_RAST, pipeline->rast.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DS, pipeline->ds.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_BLEND, pipeline->blend.state_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VS_CONST, cmd->state.shader_const_ib[MESA_SHADER_VERTEX]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_HS_CONST, cmd->state.shader_const_ib[MESA_SHADER_TESS_CTRL]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DS_CONST, cmd->state.shader_const_ib[MESA_SHADER_TESS_EVAL]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_GS_CONST, cmd->state.shader_const_ib[MESA_SHADER_GEOMETRY]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_FS_CONST, cmd->state.shader_const_ib[MESA_SHADER_FRAGMENT]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DESC_SETS, cmd->state.desc_sets_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DESC_SETS_LOAD, cmd->state.desc_sets_load_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VS_PARAMS, vs_params);
+
+ for (uint32_t i = 0; i < ARRAY_SIZE(cmd->state.dynamic_state); i++) {
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i,
+ ((pipeline->dynamic_state_mask & BIT(i)) ?
+ cmd->state.dynamic_state[i] :
+ pipeline->dynamic_state[i]));
}
+ } else {
- tu_cs_emit(cs, cp_set_draw_state);
- tu_cs_emit_qw(cs, iova);
+ /* emit draw states that were just updated
+ * note we eventually don't want to have to emit anything here
+ */
+ uint32_t draw_state_count =
+ has_tess +
+ ((cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) ? 5 : 0) +
+ ((cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) ? 1 : 0) +
+ ((cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) ? 1 : 0) +
+ 1; /* vs_params */
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_count);
+
+ /* We may need to re-emit tess consts if the current draw call is
+ * sufficiently larger than the last draw call. */
+ if (has_tess)
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_TESS, tess_consts);
+ if (cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) {
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VS_CONST, cmd->state.shader_const_ib[MESA_SHADER_VERTEX]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_HS_CONST, cmd->state.shader_const_ib[MESA_SHADER_TESS_CTRL]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DS_CONST, cmd->state.shader_const_ib[MESA_SHADER_TESS_EVAL]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_GS_CONST, cmd->state.shader_const_ib[MESA_SHADER_GEOMETRY]);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_FS_CONST, cmd->state.shader_const_ib[MESA_SHADER_FRAGMENT]);
+ }
+ if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS)
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_DESC_SETS_LOAD, cmd->state.desc_sets_load_ib);
+ if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS)
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers_ib);
+ tu_cs_emit_sds_ib(cs, TU_DRAW_STATE_VS_PARAMS, vs_params);
}
tu_cs_sanity_check(cs);
- /* track BOs */
- if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) {
- for (uint32_t i = 0; i < MAX_VBS; i++) {
- const struct tu_buffer *buf = cmd->state.vb.buffers[i];
- if (buf)
- tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
- }
- }
- if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
- unsigned i;
- for_each_bit(i, descriptors_state->valid) {
- struct tu_descriptor_set *set = descriptors_state->sets[i];
- for (unsigned j = 0; j < set->layout->buffer_count; ++j) {
- if (set->buffers[j]) {
- tu_bo_list_add(&cmd->bo_list, set->buffers[j],
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
- if (set->size > 0) {
- tu_bo_list_add(&cmd->bo_list, &set->pool->bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
- }
- }
- }
- if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS) {
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- const struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
- if (buf) {
- tu_bo_list_add(&cmd->bo_list, buf->bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
- }
-
/* There are too many graphics dirty bits to list here, so just list the
* bits to preserve instead. The only things not emitted here are
* compute-related state.
*/
- cmd->state.dirty &= TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS;
-
- /* Fragment shader state overwrites compute shader state, so flag the
- * compute pipeline for re-emit.
- */
- cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_PIPELINE;
+ cmd->state.dirty &= (TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS | TU_CMD_DIRTY_COMPUTE_PIPELINE);
return VK_SUCCESS;
}
-static void
-tu6_emit_draw_indirect(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
-{
- const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
- bool has_gs = cmd->state.pipeline->active_stages &
- VK_SHADER_STAGE_GEOMETRY_BIT;
-
- tu_cs_emit_regs(cs,
- A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
- A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
-
- if (draw->indexed) {
- const enum a4xx_index_size index_size =
- tu6_index_size(cmd->state.index_type);
- const uint32_t index_bytes =
- (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
- const struct tu_buffer *index_buf = cmd->state.index_buffer;
- unsigned max_indicies =
- (index_buf->size - cmd->state.index_offset) / index_bytes;
-
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
- CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_INDIRECT, 6);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit_qw(cs, index_buf->bo->iova + cmd->state.index_offset);
- tu_cs_emit(cs, A5XX_CP_DRAW_INDX_INDIRECT_3_MAX_INDICES(max_indicies));
- tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
- } else {
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT, 3);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
- }
-
- tu_bo_list_add(&cmd->bo_list, draw->indirect->bo, MSM_SUBMIT_BO_READ);
-}
-
-static void
-tu6_emit_draw_direct(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
-{
-
- const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
- bool has_gs = cmd->state.pipeline->active_stages &
- VK_SHADER_STAGE_GEOMETRY_BIT;
-
- tu_cs_emit_regs(cs,
- A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
- A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
-
- /* TODO hw binning */
- if (draw->indexed) {
- const enum a4xx_index_size index_size =
- tu6_index_size(cmd->state.index_type);
- const uint32_t index_bytes =
- (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
- const struct tu_buffer *buf = cmd->state.index_buffer;
- const VkDeviceSize offset = buf->bo_offset + cmd->state.index_offset +
- index_bytes * draw->first_index;
- const uint32_t size = index_bytes * draw->count;
-
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
- CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit(cs, draw->instance_count);
- tu_cs_emit(cs, draw->count);
- tu_cs_emit(cs, 0x0); /* XXX */
- tu_cs_emit_qw(cs, buf->bo->iova + offset);
- tu_cs_emit(cs, size);
- } else {
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit(cs, draw->instance_count);
- tu_cs_emit(cs, draw->count);
- }
-}
-
-static void
-tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
+static uint32_t
+tu_draw_initiator(struct tu_cmd_buffer *cmd, enum pc_di_src_sel src_sel)
{
- struct tu_cs *cs = &cmd->draw_cs;
- VkResult result;
-
- result = tu6_bind_draw_states(cmd, cs, draw);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
-
- if (draw->indirect)
- tu6_emit_draw_indirect(cmd, cs, draw);
- else
- tu6_emit_draw_direct(cmd, cs, draw);
-
- if (cmd->state.streamout_enabled) {
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- if (cmd->state.streamout_enabled & (1 << i))
- tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i, false);
- }
+ const struct tu_pipeline *pipeline = cmd->state.pipeline;
+ uint32_t initiator =
+ CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(pipeline->ia.primtype) |
+ CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(src_sel) |
+ CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(cmd->state.index_size) |
+ CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY);
+
+ if (pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT)
+ initiator |= CP_DRAW_INDX_OFFSET_0_GS_ENABLE;
+
+ switch (pipeline->tess.patch_type) {
+ case IR3_TESS_TRIANGLES:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_TRIANGLES) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
+ case IR3_TESS_ISOLINES:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_ISOLINES) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
+ case IR3_TESS_NONE:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS);
+ break;
+ case IR3_TESS_QUADS:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
}
-
- cmd->wait_for_idle = true;
-
- tu_cs_sanity_check(cs);
+ return initiator;
}
void
uint32_t firstVertex,
uint32_t firstInstance)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.count = vertexCount;
- info.instance_count = instanceCount;
- info.first_instance = firstInstance;
- info.vertex_offset = firstVertex;
+ tu6_draw_common(cmd, cs, false, firstVertex, firstInstance, vertexCount);
- tu_draw(cmd_buffer, &info);
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit(cs, vertexCount);
}
void
int32_t vertexOffset,
uint32_t firstInstance)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.indexed = true;
- info.count = indexCount;
- info.instance_count = instanceCount;
- info.first_index = firstIndex;
- info.vertex_offset = vertexOffset;
- info.first_instance = firstInstance;
+ tu6_draw_common(cmd, cs, true, vertexOffset, firstInstance, indexCount);
- tu_draw(cmd_buffer, &info);
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit(cs, indexCount);
+ tu_cs_emit(cs, 0x0); /* XXX */
+ tu_cs_emit_qw(cs, cmd->state.index_va + (firstIndex << cmd->state.index_shift));
+ tu_cs_emit(cs, indexCount << cmd->state.index_shift);
}
void
uint32_t drawCount,
uint32_t stride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.count = drawCount;
- info.indirect = buffer;
- info.indirect_offset = offset;
- info.stride = stride;
+ tu6_draw_common(cmd, cs, false, 0, 0, 0);
- tu_draw(cmd_buffer, &info);
+ for (uint32_t i = 0; i < drawCount; i++) {
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT, 3);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset + stride * i);
+ }
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
void
uint32_t drawCount,
uint32_t stride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.indexed = true;
- info.count = drawCount;
- info.indirect = buffer;
- info.indirect_offset = offset;
- info.stride = stride;
+ tu6_draw_common(cmd, cs, true, 0, 0, 0);
- tu_draw(cmd_buffer, &info);
+ for (uint32_t i = 0; i < drawCount; i++) {
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDX_INDIRECT, 6);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
+ tu_cs_emit_qw(cs, cmd->state.index_va);
+ tu_cs_emit(cs, A5XX_CP_DRAW_INDX_INDIRECT_3_MAX_INDICES(cmd->state.max_index_count));
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset + stride * i);
+ }
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
void tu_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer,
uint32_t counterOffset,
uint32_t vertexStride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _counterBuffer);
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _counterBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- struct tu_draw_info info = {};
+ tu6_draw_common(cmd, cs, false, 0, firstInstance, 0);
- info.instance_count = instanceCount;
- info.first_instance = firstInstance;
- info.streamout_buffer = buffer;
- info.streamout_buffer_offset = counterBufferOffset;
- info.stride = vertexStride;
+ tu_cs_emit_pkt7(cs, CP_DRAW_AUTO, 6);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_XFB));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + counterBufferOffset);
+ tu_cs_emit(cs, counterOffset);
+ tu_cs_emit(cs, vertexStride);
- tu_draw(cmd_buffer, &info);
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
struct tu_dispatch_info
struct tu_pipeline *pipeline = cmd->state.compute_pipeline;
struct tu_descriptor_state *descriptors_state =
&cmd->descriptors[VK_PIPELINE_BIND_POINT_COMPUTE];
- VkResult result;
+
+ /* TODO: We could probably flush less if we add a compute_flush_bits
+ * bitfield.
+ */
+ tu_emit_cache_flush(cmd, cs);
if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_PIPELINE)
tu_cs_emit_ib(cs, &pipeline->program.state_ib);
tu_emit_compute_driver_params(cs, pipeline, info);
- if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS) {
- result = tu6_emit_descriptor_sets(cmd, pipeline,
- VK_PIPELINE_BIND_POINT_COMPUTE, &ib,
- false);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
-
- /* track BOs */
- unsigned i;
- for_each_bit(i, descriptors_state->valid) {
- struct tu_descriptor_set *set = descriptors_state->sets[i];
- for (unsigned j = 0; j < set->layout->buffer_count; ++j) {
- if (set->buffers[j]) {
- tu_bo_list_add(&cmd->bo_list, set->buffers[j],
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
-
- if (set->size > 0) {
- tu_bo_list_add(&cmd->bo_list, &set->pool->bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
- }
- }
- }
-
- if (ib.size)
- tu_cs_emit_ib(cs, &ib);
-
- if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS)
+ if ((cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS) &&
+ pipeline->load_state.state_ib.size > 0) {
tu_cs_emit_ib(cs, &pipeline->load_state.state_ib);
+ }
cmd->state.dirty &=
~(TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS | TU_CMD_DIRTY_COMPUTE_PIPELINE);
- /* Compute shader state overwrites fragment shader state, so we flag the
- * graphics pipeline for re-emit.
- */
- cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
-
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_COMPUTE));
}
tu_cs_emit_wfi(cs);
-
- tu6_emit_cache_flush(cmd, cs);
}
void
tu_cs_discard_entries(&cmd_buffer->draw_epilogue_cs);
tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
+ cmd_buffer->state.cache.pending_flush_bits |=
+ cmd_buffer->state.renderpass_cache.pending_flush_bits;
+ tu_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier, true);
+
cmd_buffer->state.pass = NULL;
cmd_buffer->state.subpass = NULL;
cmd_buffer->state.framebuffer = NULL;
const VkImageMemoryBarrier *pImageMemoryBarriers,
const struct tu_barrier_info *info)
{
- /* renderpass case is only for subpass self-dependencies
- * which means syncing the render output with texture cache
- * note: only the CACHE_INVALIDATE is needed in GMEM mode
- * and in sysmem mode we might not need either color/depth flush
+ struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
+ VkAccessFlags srcAccessMask = 0;
+ VkAccessFlags dstAccessMask = 0;
+
+ for (uint32_t i = 0; i < memoryBarrierCount; i++) {
+ srcAccessMask |= pMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pMemoryBarriers[i].dstAccessMask;
+ }
+
+ for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
+ srcAccessMask |= pBufferMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pBufferMemoryBarriers[i].dstAccessMask;
+ }
+
+ enum tu_cmd_access_mask src_flags = 0;
+ enum tu_cmd_access_mask dst_flags = 0;
+
+ for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
+ TU_FROM_HANDLE(tu_image, image, pImageMemoryBarriers[i].image);
+ VkImageLayout old_layout = pImageMemoryBarriers[i].oldLayout;
+ /* For non-linear images, PREINITIALIZED is the same as UNDEFINED */
+ if (old_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
+ (image->tiling != VK_IMAGE_TILING_LINEAR &&
+ old_layout == VK_IMAGE_LAYOUT_PREINITIALIZED)) {
+ /* The underlying memory for this image may have been used earlier
+ * within the same queue submission for a different image, which
+ * means that there may be old, stale cache entries which are in the
+ * "wrong" location, which could cause problems later after writing
+ * to the image. We don't want these entries being flushed later and
+ * overwriting the actual image, so we need to flush the CCU.
+ */
+ src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ }
+ srcAccessMask |= pImageMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pImageMemoryBarriers[i].dstAccessMask;
+ }
+
+ /* Inside a renderpass, we don't know yet whether we'll be using sysmem
+ * so we have to use the sysmem flushes.
*/
- if (cmd->state.pass) {
- tu6_emit_event_write(cmd, &cmd->draw_cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, &cmd->draw_cs, PC_CCU_FLUSH_DEPTH_TS, true);
- tu6_emit_event_write(cmd, &cmd->draw_cs, CACHE_INVALIDATE, false);
- return;
+ bool gmem = cmd->state.ccu_state == TU_CMD_CCU_GMEM &&
+ !cmd->state.pass;
+ src_flags |= vk2tu_access(srcAccessMask, gmem);
+ dst_flags |= vk2tu_access(dstAccessMask, gmem);
+
+ struct tu_cache_state *cache =
+ cmd->state.pass ? &cmd->state.renderpass_cache : &cmd->state.cache;
+ tu_flush_for_access(cache, src_flags, dst_flags);
+
+ for (uint32_t i = 0; i < info->eventCount; i++) {
+ TU_FROM_HANDLE(tu_event, event, info->pEvents[i]);
+
+ tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_READ);
+
+ tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
+ CP_WAIT_REG_MEM_0_POLL_MEMORY);
+ tu_cs_emit_qw(cs, event->bo.iova); /* POLL_ADDR_LO/HI */
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
}
}
}
static void
-write_event(struct tu_cmd_buffer *cmd, struct tu_event *event, unsigned value)
+write_event(struct tu_cmd_buffer *cmd, struct tu_event *event,
+ VkPipelineStageFlags stageMask, unsigned value)
{
struct tu_cs *cs = &cmd->cs;
- tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_WRITE);
+ /* vkCmdSetEvent/vkCmdResetEvent cannot be called inside a render pass */
+ assert(!cmd->state.pass);
- /* TODO: any flush required before/after ? */
+ tu_emit_cache_flush(cmd, cs);
- tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
- tu_cs_emit_qw(cs, event->bo.iova); /* ADDR_LO/HI */
- tu_cs_emit(cs, value);
+ tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_WRITE);
+
+ /* Flags that only require a top-of-pipe event. DrawIndirect parameters are
+ * read by the CP, so the draw indirect stage counts as top-of-pipe too.
+ */
+ VkPipelineStageFlags top_of_pipe_flags =
+ VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT |
+ VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT;
+
+ if (!(stageMask & ~top_of_pipe_flags)) {
+ tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
+ tu_cs_emit_qw(cs, event->bo.iova); /* ADDR_LO/HI */
+ tu_cs_emit(cs, value);
+ } else {
+ /* Use a RB_DONE_TS event to wait for everything to complete. */
+ tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 4);
+ tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(RB_DONE_TS));
+ tu_cs_emit_qw(cs, event->bo.iova);
+ tu_cs_emit(cs, value);
+ }
}
void
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_event, event, _event);
- write_event(cmd, event, 1);
+ write_event(cmd, event, stageMask, 1);
}
void
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_event, event, _event);
- write_event(cmd, event, 0);
+ write_event(cmd, event, stageMask, 0);
}
void
const VkImageMemoryBarrier *pImageMemoryBarriers)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *cs = &cmd->cs;
-
- /* TODO: any flush required before/after? (CP_WAIT_FOR_ME?) */
-
- for (uint32_t i = 0; i < eventCount; i++) {
- TU_FROM_HANDLE(tu_event, event, pEvents[i]);
+ struct tu_barrier_info info;
- tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_READ);
+ info.eventCount = eventCount;
+ info.pEvents = pEvents;
+ info.srcStageMask = 0;
- tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
- CP_WAIT_REG_MEM_0_POLL_MEMORY);
- tu_cs_emit_qw(cs, event->bo.iova); /* POLL_ADDR_LO/HI */
- tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
- }
+ tu_barrier(cmd, memoryBarrierCount, pMemoryBarriers,
+ bufferMemoryBarrierCount, pBufferMemoryBarriers,
+ imageMemoryBarrierCount, pImageMemoryBarriers, &info);
}
void